Patients enter through an Agentforce-driven intake assistant. Structured signals are then prioritized for women 40-60 using symptom clusters, endocrine context, and safety-first clinical markers.
If you're here because something feels off — the sleep, the mood, the symptoms you can't quite explain — you're not imagining it, and you're not alone. This intake gathers your whole story so your care team can point you to the right next step sooner. Advisory and synthetic in this prototype — a clinician always makes the final call.
Loading intake stage…
Another way in
Prefer to talk? Voice is the next channel.
Agentforce Voice routes to the same intake agent as the chat above — same subagents, same Data 360 grounding, same Care Router handoff. Its state below reflects what this deployment can actually prove.
Upstream of intake
The acquisition funnel that feeds this agent
Before a patient reaches the intake agent above, a lead flows through three Agentforce agents over Google A2A: Inbound Lead Generation → Qualification → then either Patient Intake → Care Router (qualified & ready) or Prospecting & Nurture (warming). Each hop enforces its own Agent Fabric policies. Run a scenario, then open the trace.
Validated-instrument scoring
The Assessment agent that grades intake severity
The Assessment agent administers a validated instrument (MRS, Greene, PHQ-9, ISI) over Google A2A and scores it deterministically — real cutoff math, no LLM. The score maps onto the intake severity the Care Router consumes, and every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Eligibility & benefit verification (EBV)
The Benefits agent that verifies coverage before care
The Benefits & Coverage Verification agent runs a deterministic synthetic EBV round-trip over Google A2A — plan status, in/out-of-network, deductible + amount met, coinsurance/copay, and an estimated visit cost + patient responsibility, each tracing to a (mock) payer/clearinghouse source. This is a labeled demo mock — not a live 270/271 or FHIR eligibility call. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Appointment scheduling (book / reschedule)
The Scheduling agent that books the recommended visit
The Appointment Scheduling agent books (or reschedules) the MSCP menopause-specialist visit the Care Router recommends over Google A2A, honoring the requested modality against a deterministic synthetic provider calendar and returning a confirmed slot with a synthetic ServiceAppointment id + source provenance, then handing the booking to the Engagement Agent for reminders. This is a labeled demo mock — not a real Salesforce Scheduler / ServiceAppointment write. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Care-plan instantiation & summary (live Claude)
The Care Plan agent — the second live-Claude agent
The Care Plan agent deterministically instantiates a menopause care plan from a defined template (goals, interventions, follow-up cadence) based on the Care Router pathway/severity + intake, then writes a non-prescriptive progress summary with live Anthropic Claude, falling back to a deterministic scripted summary (with a recorded reason) on a missing key or any SDK error — just like the Care Router. It is governed by the same model allow-list and never commits a clinical action without a clinician. The templates are illustrative synthetics, not a certified care-plan engine. Run a preset, then open the trace.
The Clinical Summary agent — the third live-Claude agent
The Clinical Summary agent composes the outputs the other agents already produced (intake, the Care Router pathway, and any care plan) into two artifacts — a patient after-visit summary and a clinician handoff. The context is assembled deterministically from only the facts that are present, so the summary can only assert what the upstream agents established — the grounding guarantee is real. The phrasing is written with live Anthropic Claude, falling back to a deterministic scripted composition (with a recorded reason) on a missing key or any SDK error — just like the Care Plan agent. It is governed by the same model allow-list and commits no clinical action. The artifacts are illustrative synthetics, not a certified clinical-documentation engine. Run a preset, then open the trace.
Preset scenarios
Whole-person care · social needs
The SDOH agent that screens social needs and drafts referrals
The SDOH Screening agent deterministically screens the patient with a validated instrument (the CMS AHC-HRSN core domains: housing, food, transportation, utilities, interpersonal safety), flags the positive social-need domains, escalates a positive interpersonal-safety screen to a human social worker, and drafts consent-gated community-resource referrals (211, food bank, housing/utility assistance, a domestic-violence hotline) — human-approval-gated, never an autonomous enrollment. The community-resource catalog is illustrative synthetic, not a live directory of real programs. SDOH is separate from clinical severity — it raises a care-coordination flag. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Patient education & health coaching (live Claude)
The Patient Education agent — the fourth live-Claude agent
The Patient Education agent deterministically selects education modules from a defined evidence-sourced catalog (bone health, cardiovascular risk, sleep hygiene, vasomotor self-management, mood/stress, nutrition, physical activity) based on the intake symptoms/severity + upstream Care Plan focus areas + detected care gaps, then writes a warm, motivational coaching message with live Anthropic Claude, falling back to a deterministic scripted message (with a recorded reason) on a missing key or any SDK error — just like the Care Plan agent. It is general education only (no diagnosis, dosing, or individualized medical advice), every module traces to a defined evidence source, and any coaching outreach is consent-gated and human-approval-gated. The modules + source labels are illustrative synthetics, not a certified patient-education engine. Run a preset, then open the trace.
Preset scenarios
Proactive care-gap closure
The Care Gap agent that closes preventive-care gaps
The Care Gap Closure agent grounds on the patient's Data 360 context and deterministically detects menopause-relevant preventive-care gaps (bone-density/DEXA, lipid panel, mammogram, HRT follow-up) against an explicit as-of date, then drafts consent- and quiet-hours-aware outreach — human-approval-gated, never auto-sent — and hands it to the Engagement Agent. Every gap references a defined clinical-measure catalog id. The measures + intervals are illustrative synthetics, not a certified guideline engine. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
The agent that tracks longitudinal trends and routes them to a clinician
The Remote Patient Monitoring agent ingests longitudinal, self-reported and wearable/device readings (hot-flash frequency, sleep, mood, resting heart rate, weight) and deterministically detects each metric's trend (improving / stable / worsening) over the reading window, applying synthetic red-flag thresholds. Worsening or red-flag trends are routed to a human clinician for review — the agent never takes an autonomous clinical action. The metrics + thresholds are illustrative synthetics, not a certified remote-monitoring device. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Population health & risk stratification
The agent that stratifies a whole panel and prioritizes outreach
The Population Health agent reasons over a whole panel of patients at once, taking already-produced per-patient signals (intake severity, assessment band, care gaps, SDOH domains, medication adherence, monitored trend) and deterministically scoring each patient with a transparent, additive risk model into a risk tier (low / rising / high), then building a prioritized outreach worklist for a human care manager. Every tier is explainable by its contributing factors, the model uses no protected-class attributes, and a tier never triggers an autonomous care decision. The factors, weights, cutoffs, and patient references are illustrative synthetics, not a certified risk-stratification model. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Consent & preferences management
The authoritative consent ledger the other agents' consent gates defer to
The Consent & Preferences Management agent is the source of truth for consent — it holds, per patient, a consent ledger (scopes, each granted / withheld / revoked with a recorded basis and optional expiry) and communication preferences (allowed channels, quiet hours, preferred language, frequency cap), and answers one deterministic question: may this patient be contacted / have data used for this scope over this channel at this time? Every consent state traces to a recorded basis, a revocation or expiry is honored immediately, and a decision never overrides a scope. The scopes, sources, and preferences are illustrative synthetics, not a certified consent-management system. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Clinical trials & research matching
The agent that matches a patient to research studies — consent-gated, never auto-enrolled
The Clinical Trials agent matches a single patient against a synthetic study catalog using structured eligibility criteria (age band, symptom profile, comorbidities, geography, prior therapy), returns the matching studies ranked with per-criterion explanations, and drafts a consent-gated outreach. Every eligibility determination traces to a defined criterion, outreach is gated on the patient's research consent (it defers to the Consent & Preferences Management agent's research scope), and the agent never auto-enrolls — enrollment requires informed consent and a human. The study catalog, sponsors, and criteria are illustrative synthetics, not real studies or a certified eligibility engine. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Language access & health equity
The agent that ensures LEP patients understand their care — qualified interpreters, approved translations, no machine-translated consent
The Language Access agent determines a patient's preferred language (deferring to the Consent & Preferences Management agent's preferred-language preference), decides whether a qualified medical interpreter is needed and of which modality, checks whether needed materials are available in that language from an approved translated-materials catalog, and flags equity gaps. Clinical interpretation uses a qualified medical interpreter only (never a family / ad-hoc / machine interpreter), in-language materials trace to an approved source, and machine translation is never used for clinical consent. When no qualified interpreter is available it escalates to a human coordinator — never an unqualified fallback. The languages, interpreter availability, materials, and provenance are illustrative synthetics, not a certified language-access system. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
HEDIS & quality reporting
The agent that rolls a panel of patients into HEDIS quality-measure compliance — never autonomously submitted
The HEDIS agent ingests already-produced per-patient signals across a panel and deterministically rolls them up into HEDIS quality-measure compliance — numerator, denominator, catalog-sourced exclusions, and compliance rate per measure — the artifact provider organizations owe payers under value-based-care contracts. Every measure must trace to the defined HEDIS measure catalog, every applied denominator exclusion must trace to a defined catalog exclusion on that measure, and every submission package requires human quality-team approval — the agent NEVER autonomously files to a payer / CMS / quality registry. The HEDIS measure catalog, denominator windows, numerator thresholds, and exclusion lists are illustrative synthetics, not NCQA-certified specifications. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Advance care planning (midlife touchpoint)
The agent that surfaces a patient's advance directives — never applies a change autonomously
The ACP agent uses perimenopause / menopause as a natural midlife touchpoint to surface which advance directives are on file (living will, DPOA-HC; POLST only for serious-illness), flag missing / stale / language-access gaps, and draft a consent-gated conversation prompt for the care team to deliver. Every directive on file must trace to the catalog + an approved source, every directive change is clinician + patient sign-off gated — the agent NEVER autonomously creates, updates, or overrides a directive — and for a limited-English-proficiency (LEP) patient the active prompt is withheld until a qualified-interpreter plan is documented (a safe answer, not a governance block). The directive catalog, source labels, and staleness threshold are illustrative synthetics, not a certified advance-directives registry. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Care team & case management
The agent that coordinates the multi-disciplinary team around a high-need patient — never a roster edit without the case manager
The care-team agent assembles the multi-disciplinary team around a single high-need menopause/midlife patient — PCP, MSCP, cardiology, endocrinology, bone-health, pelvic-floor PT, behavioral health — deterministically resolves which roles are needed from the patient's active clinical needs, assigns a case manager by a stable hash on the patient ref, and emits a shared team snapshot for the whole team. Every role must trace to the care-role catalog, every roster change is case-manager sign-off gated (no autonomous add/remove), and a legitimate team must include a PCP anchor. The care-role catalog, condition→role triggers, case-manager pool, and refs are illustrative synthetics, not a certified care-team schema. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Discharge & transitions of care
The agent that closes the loop back to primary care after a hospitalization — never a “recommended” follow-up, never an autonomous med change
The TOC agent runs the close-the-loop workflow after a hospitalization / ED / observation encounter — it reconciles the discharge medication list (added, removed, dose-changed, unchanged), books the follow-up (or hands off to the Appointment Scheduling agent — never a text recommendation), pulls the encounter-reason red-flag warning signs, emits the teach-back checklist, and assembles the PCP handoff summary. Every medication must trace to an approved source, every medication change is clinician-signoff gated (no autonomous changes), and the follow-up is a scheduled slot or an explicit awaiting-schedule handoff — the load-bearing 30-day- readmission guard against “recommended” follow-ups masquerading as complete. The encounter categories, red-flag catalog, follow-up window, approved-source labels, and teach-back items are illustrative synthetics, not a certified TOC system. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Grievance & appeals
The agent that classifies member complaints and coverage denials — never resolves them, never leaks PHI, never extends a regulatory deadline
The grievance-and-appeals agent runs the intake half of the regulated process — classifying a member complaint or coverage-denial appeal (grievance / billing / standard-appeal / expedited-appeal), routing it to the correct human queue (member-services / clinical-review / compliance), and stamping a regulatory deadline that traces to the case-type catalog + received date. It NEVER resolves, approves, or denies a case on its own; every case is queued for human review. The routing summary handed to the receiving queue is PHI-safe (structured only — no free-text PHI), so it can be delivered via lower-trust channels (Slack, email, ticketing) without leaking PHI. The case-type catalog, deadline windows, and queue mapping are illustrative synthetics, not Medicare Advantage Chapter 13 or a real appeal-adjudication engine. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Provider credentialing & directory
The agent that fixes the ghost network — no referral to an expired / sanctioned provider, no stale directory response returned as authoritative
The credentialing agent sits alongside the data substrate and gates every referral / scheduling attempt at the network boundary. It verifies each credential (state license, DEA, board cert, sanctions clearance, NPI) against approved sources (state-medical-board, DEA-registry, ABMS-board, OIG-LEIE-sanctions, NPI-registry), computes the No-Surprises-Act freshness flag (90-day accuracy window), and emits gate flags (canReferPatient / canBookAppointment / canReturnInDirectoryResponse) the Referral Management, Appointment Scheduling, and Transitions of Care agents can consult before handing off. Sanctioned status has highest precedence — a sanctioned provider never slips through, even when other credentials look complete. The catalog, verification sources, NSA window, and directory schema are illustrative synthetics, not NCQA / CAQH credentialing or a live state-medical-board / OIG-LEIE feed. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Quality-measure attribution
The other half of the HEDIS story — who gets the credit / accountability for the rate
The attribution agent pairs with the HEDIS & Quality Reporting agent: HEDIS computes the rates, this agent decides whose panel each patient counts on. It attributes every patient to a provider / clinic / VBC contract under a defined methodology from the catalog (plurality-of-visits, PCP-of-record, prospective Medicare Advantage, contract-defined window), honors the contract's exclusion terms (age band, network status, exclusion codes) so the scorecard isn't polluted with excluded patients, and applies a documented tie-break chain (most-recent-visit-wins then provider-ref-lexical-ascending) when the primary metric ties — no coin-flip, no gameable non-determinism. Rolls up per-provider counts so downstream HEDIS scoring lands on the right denominator. The methodology catalog, contract catalog, tie-break rules, and refs are illustrative synthetics, not CMS Shared Savings / ACO REACH / NCQA attribution. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Complex Care Management (CCM) · reimbursable time-tracking
The agent that assembles the CPT 99490 / 99491 / 99487 / 99489 billing package — never autonomously submits to CMS
The CCM agent runs the reimbursable time-tracking piece of care management for a Medicare-eligible high-need patient — it confirms CCM eligibility (≥ 2 catalog-sourced chronic conditions, Medicare age, coverage flag, consent), tracks per-activity monthly minutes against a defined activity catalog, maps the total to the CPT ladder (99490 → 99491 → 99487 → 99489), and assembles a billing package for human quality-team review. It NEVER autonomously submits a CMS claim, and every logged minute traces to a catalog activity + sums to the reported total (the guard against phantom-minute inflation, the classic CCM audit finding). The chronic-condition catalog, CCM activity catalog, CPT thresholds, and Medicare flags are illustrative synthetics, not CMS Chapter 12 CCM billing. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
The agent that clean-pays, pends, or drafts a denial with a specific reason code — never autonomously finalizes a denial
The claims-adjudication agent applies payer-specific catalog edits (NCCI-PTP unbundling, LCD/NCD coverage, benefit limits, prior-auth linkage, duplicates, network, timely-filing) to each submitted claim, classifies as clean-pay / pend / deny-drafted with a specific catalog reason code, and routes anything non-clean to a human. The agent NEVER autonomously finalizes a denial — every denial is DRAFTED for adjudicator cosign, because denial letters are legally consequential under CMS / ERISA / state insurance code. The edit catalog, reason-code catalog, and benefit rules are illustrative synthetics, not CMS X12 837 / NCCI PTP / LCD/NCD or real payer benefit configuration. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Formulary & Drug Utilization Review · first-pass payer-side
The agent that preferred-approves, pends for step-therapy / quantity / interaction / non-formulary, and never autonomously overrides an exception
The formulary agent evaluates a proposed medication against the payer's catalog rules (tier, step-therapy, quantity limits, drug-drug interactions), classifies as preferred-approved / pend with a specific catalog reason code, and routes pends to a clinician (or pharmacist for interactions). The agent NEVER autonomously overrides a formulary exception — every non-preferred decision is DRAFTED for clinician cosign, because formulary exceptions require a prescriber's documented rationale under Medicare Advantage Chapter 6 + Part D. Menopause-relevant because HRT tier placement varies significantly by plan (transdermal estradiol is often Tier 2 or non-formulary despite being clinically preferred for CVD-risk profiles). The drug catalog, rule catalog, step-therapy chains, and interaction pairs are illustrative synthetics, not Medi-Span / RxNorm / a real payer formulary. Every run is governed by the Agent Fabric.
The agent that flags suspicious patterns — never denies a claim, never freezes payment, never scores on protected-class factors
The FWA agent screens claims against catalog-defined patterns (unbundling, upcoding, duplicate billing, quantity outliers, impossible-day billing, phantom services), classifies each hit by severity, and routes to the SIU for HUMAN review. It NEVER autonomously denies a claim, opens an investigation, or freezes payment — those are formal acts under Section 1557 / state insurance code / due process. The engine may NOT score on protected-class attributes (a well-documented compliance failure in real payer FWA systems). Distinct from Claims Adjudication (which AUTO-denies mechanical edits with a reason code): FWA is about suspicious patterns that need investigation. The pattern catalog, peer baselines, and severity thresholds are illustrative synthetics, not SAS / LexisNexis / a real payer SIU rule set. Every run is governed by the Agent Fabric.
The agent that pays per-visit stipends against IRB-approved schedules — never deviates without coordinator cosign, never pays without consent
The trial-payments agent pairs with Clinical Trials Matching. For each participant visit it looks up the IRB-approved compensation schedule (trial + visit type + IRB approval ref), verifies research-payment informed consent is on file (45 CFR 46 requirement), computes the stipend + travel reimbursement, and routes non-standard payments (missed visit, out-of-range travel, extra procedure) to the study coordinator for cosign. The agent NEVER autonomously deviates from an IRB-approved schedule and NEVER pays a non-consented participant. The trial catalog, IRB schedules, visit types, rules, and travel rates are illustrative synthetics, not IRBNet / WCG IRB / Advarra IRB or a real sponsor's payment protocol. Every run is governed by the Agent Fabric.
Preset scenarios
Utilization Review · MCG/InterQual analog · pre-service medical necessity
The agent that screens a proposed procedure or admission against catalog criteria — never autonomously denies, every SLA traces to catalog
The utilization-review agent runs the pre-service medical-necessity screen for a proposed procedure or inpatient admission against the catalog criteria set for that service type, classifies as approves-meets-criteria / pend-for-clinical-review / require-peer-to-peer / blocked-non-covered, and routes non-approved cases to a clinical reviewer or peer-to-peer with a catalog-sourced SLA deadline (standard 72h, urgent 24h, concurrent-review 24h). Distinct from Prior Authorization (assembly) and Claims Adjudication (post-service edits). The agent NEVER autonomously denies — every non-approved decision is DRAFTED for clinician cosign (Medicare Advantage / state UR-agent codes require notice + due-process rights). The service-type catalog, criteria sets, rules, reason codes, and SLA windows are illustrative synthetics, not MCG (Milliman Care Guidelines / Indicia), InterQual, or a real payer's UR rule set. Every run is governed by the Agent Fabric.
The agent that classifies provider-network contracts and computes VBC benchmarks — never autonomously commits a term change
The provider-contracting agent runs on the commercial plane (no PHI) alongside the Pipeline Management and Account Management agents. For each provider-network contract it classifies the payment model (FFS, capitation, shared-savings, bundled-payment, MA-VBC, commercial-VBC), computes the quality-gate + spend-benchmark drift for a caller- provided reporting period against a catalog methodology, and classifies as in-good-standing / benchmark-drift-review / draft-term-change / blocked-non-catalog-contract. The agent NEVER autonomously commits a contract-term change — every draft is DRAFTED for account-owner cosign (state insurance code / provider-contract law / CMS Medicare Advantage require a human owner sign-off). The contract-type catalog, methodology catalog, rules, and reason codes are illustrative synthetics, not Salesforce Health Cloud Provider Network Management, Optum Contract Manager, or a real payer's contract-lifecycle system. Every run is governed by the Agent Fabric.
Deal Desk — a bounded discount, never an autonomous out-of-guardrail approval
Pause’s own go-to-market tooling (no patient PHI). Given a proposed quote (line items each a product, list price, quantity, and proposed discount %), this agent deterministically prices each line, sums the totals, computes the effective blended discount, and checks each line’s discount against its product’s guardrail. A standard quote auto-approves; any line over guardrail escalates to a human deal-desk owner — it never autonomously approves an out-of-guardrail discount. The product catalog and guardrails are illustrative, not a certified CPQ system. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Care Coordination Handoff · cross-setting SBAR (Joint Commission NPSG-2)
The agent that assembles the SBAR handoff for any cross-setting patient transition — never routes to an uncredentialed clinician, never discloses PHI without consent
The handoff agent handles any cross-setting patient transition (hospital → SNF, SNF → home, home → hospice, ED → PCP, PCP → specialist, PCP → behavioral health). For each transition it assembles the Joint-Commission-NPSG-2 SBAR (situation, background, assessment, recommendation), verifies the receiving clinician's credentialing status, and confirms transfer consent for transitions that share PHI with a new setting. Distinct from Transitions of Care (post-discharge hospital→home + med reconciliation) and Referral Management (outbound specialist referral). The agent NEVER autonomously accepts on behalf of the receiving clinician; every accepted handoff is DRAFTED for the receiving clinician's cosign. The care-setting catalog, transition-type catalog, SBAR rule set, and reason codes are illustrative synthetics, not Epic Care Everywhere, Cerner CareAware, or a real health system's handoff protocol. Every run is governed by the Agent Fabric.
Preset scenarios
Adverse Event Reporting · FDA MedWatch / VAERS analog
The agent that classifies drug ADRs / vaccine reactions / device malfunctions and drafts MedWatch or VAERS — never autonomously files to the FDA
The adverse-event agent runs a pharmacovigilance / device-safety reporting pipeline. For each reported event (drug ADR, vaccine reaction, device malfunction, medication error, therapeutic failure) it computes the 21-CFR-314.80 seriousness tier (non-serious / serious / life-threatening / death) from caller-provided outcome flags, verifies reporter identity attestation, and classifies as draft-medwatch (3500 / 3500A) / draft-vaers / blocked-non-catalog-event / blocked-reporter-unverified. All drafts route to a regulatory-team queue for cosign. The agent NEVER autonomously files to the FDA (21 CFR 314.80 mandatory reporting has sponsor / manufacturer / clinician liability), and NEVER drafts on an unverified reporter (FDA reporting requires an attested reporter). The event-type catalog, seriousness tiers, rules, and reason codes are illustrative synthetics, not FDA MedWatch, VAERS, EudraVigilance, or a real sponsor's pharmacovigilance database. Every run is governed by the Agent Fabric.
The agent that classifies cross-org PHI exchanges (TEFCA / Carequality / CommonWell) — never releases non-TPO PHI without consent, never releases to an unverified participant
The data-sharing agent handles cross-organization PHI exchanges over TEFCA QHIN / Carequality / CommonWell / Direct Secure Messaging. For each request it classifies the exchange purpose (treatment / payment / operations / patient-request / public-health / research), verifies the counterparty is a Trusted Exchange Framework participant, applies the patient's data-sharing consent scopes from the Consent agent, and classifies as release-authorized / pend-purpose-verification / blocked-non-catalog-purpose / blocked-participant-unverified / blocked-consent-required-non-tpo. The agent NEVER autonomously releases PHI for a non-TPO purpose without an active consent scope (HIPAA §164.506 boundary — TPO doesn't need consent, everything else does) and NEVER releases to an unverified counterparty (45 CFR 171 / TEFCA Common Agreement). The exchange-network catalog, exchange-purpose catalog, rules, and reason codes are illustrative synthetics, not an actual TEFCA QHIN implementation, the Carequality Interoperability Framework, or a certified ONC data-sharing gateway. Every run is governed by the Agent Fabric.
Preset scenarios
Risk adjustment & HCC coding
The agent that finds documented-but-uncoded conditions — evidence-supported HCCs, a RAF-style score, clinician-validated, never autonomously submitted
The Risk Adjustment agent reviews a patient's clinical context and identifies suspected / confirmed HCCs (Hierarchical Condition Categories) for value-based care, mapping each to the documented clinical evidence that supports it, computing a RAF-style risk score from the confirmed set, and flagging coding gaps (evidence documented but uncoded) and unsupported / over-coded entries (coded but not documented). It COMPLEMENTS — it does not duplicate — the quality agents: those score quality measures; this is risk-adjustment condition coding. Every confirmed / suspected HCC must trace to documented clinical evidence (no upcoding), every suspected code is a recommendation requiring clinician validation, and the agent never autonomously submits codes or adjusts a claim / RAF. A coding gap and an unsupported flag are safe, honest outputs surfaced for a clinician — not blocks. The HCC catalog, RAF weights, and evidence are illustrative synthetics, not the certified CMS-HCC model, real RAF coefficients, or a certified coding engine. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Master patient index / identity resolution
The identity/dedup layer — transparent demographic matching, human-review-gated merges, never a protected-class feature
The Master Patient Index agent resolves a patient's identity across source systems: given an incoming record plus a set of candidate records, it deterministically scores each candidate against a transparent weighted feature set (name, DOB, member/MRN identifier, address, phone, administrative sex), classifies each as match / possible-match / no-match by fixed thresholds, and recommends link / merge / manual-review / no-action. It is a recommender + integrity gate: a high-confidence match surfaces a link/merge, but a merge below the auto-match threshold requires a human steward (there is never an auto-merged state), it never autonomously merges a low-confidence pair, and it never uses a protected-class attribute as a matching feature. The match features, weights, thresholds, and records are illustrative synthetics, not a certified EMPI algorithm. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
The Break-the-Glass agent governs emergency override access to PHI: given an access request (requester role, target patient, stated purpose, an emergency flag, and a clinical justification), it deterministically decides whether to grant access. A grant is always time-boxed (an expiry derived from the request's own time) and minimum-necessary (a scoped field set — never the full chart), always emits a mandatory audit event, and is flagged for mandatory post-access review. It never grants standing / broad / full-record access and never grants without a recorded justification. A deny (no emergency, no justification, or an off-catalog purpose) is a safe, completed answer — not a block. The purpose catalog, scopes, durations, and audit ids are illustrative synthetics, not a certified break-the-glass system. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Data retention & records lifecycle management
Records disposition — legal-hold-overrides-purge, schedule-sourced, never an autonomous purge
The Data Retention agent governs the lifecycle of records against retention schedules and legal holds: given a record (type/category, patient, created / last-touched dates, jurisdiction, and any active legal hold), it deterministically produces a recommendation — retain, eligible-for-purge, or hold — citing the governing retention rule and the computed expiry. It never autonomously purges: an eligible-for-purge is a recommendation requiring human approval, and an active legal hold always overrides a purge (a held record is never marked eligible-for-purge). An eligible-for-purge recommendation is a safe, completed answer — not a block, and never a deletion. The retention schedules, periods, and rule ids are illustrative synthetics, not a certified records-management system — real retention is jurisdiction-specific and legally reviewed. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
De-identification · HIPAA Safe Harbor · data substrate
De-identification — all eighteen categories screened, a recognized method cited, never a re-identifiable release
Given a dataset’s fields, the De-Identification agent deterministically screens them against the eighteen HIPAA Safe Harbor identifier categories (45 CFR 164.514(b)(2)), computes which categories still remain identifiable, and decides whether the dataset is de-identified — only if all eighteen categories are screened, a recognized method is cited, and no identifier remains. A re-identifiable dataset is never released as de-identified; it requires human review under a data use agreement. The Safe Harbor category catalog and generalization rules are illustrative synthetics, not a certified de-identification engine — a real determination applies the full Safe Harbor method or a qualified Expert Determination under 45 CFR 164.514(b). Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Minimum necessary · purpose-of-use scoping · data substrate
Minimum Necessary — purpose-of-use-sourced, scoped release, never an autonomous over-disclosure
Given a disclosure request (requestor role, purpose-of-use, requested fields by category, record scope), the Minimum Necessary agent deterministically resolves the governing purpose-of-use rule and decides per field whether it is within the minimum-necessary scope (release) or beyond it (withhold). Treatment is exempt; an over-scope or bulk disclosure requires human review, never an autonomous release. The purpose catalog and category mappings are illustrative synthetics, not a certified minimum-necessary engine — a real determination uses the covered entity's role-based access policies under 45 CFR 164.502(b) / 164.514(d). Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Audit log integrity · tamper-evidence · data substrate
Audit Log Integrity — verified only over an intact chain, never an autonomous redaction
Every agent on the fabric writes a HIPAA audit span; this agent verifies the audit trail itself. Given an audit log (entries chained by hash, with sequence numbers), it deterministically recomputes the hash chain and checks the sequence for gaps, deciding whether the log is verified. A break is flagged for human forensic review — the log is never rewritten or repaired. The hash is an illustrative non-cryptographic FNV-1a, not a certified tamper-evidence system — a real control uses SHA-256, append-only / WORM storage, and signed checkpoints. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Accounting of disclosures · HIPAA §164.528 · data substrate
Accounting of Disclosures — every accountable disclosure, never an autonomous suppression
Given a patient’s disclosure log (each disclosure a date, a recipient, and the cited purpose-of-disclosure), an as-of date, and a lookback window, this agent deterministically classifies each disclosure against the §164.528 rules (treatment / payment / operations and patient-authorized are excluded; non-TPO disclosures are accountable), filters to the window, and assembles the accounting. It classifies and assembles — it never deletes or suppresses a logged disclosure, and the accounting is a recommendation requiring privacy-officer review. The purpose catalog and accountability rules are illustrative, not a certified §164.528 system. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Right of access · HIPAA §164.524 · platform / data substrate
Right of Access — a computed deadline, never an autonomous release or denial
A patient’s §164.524 right to GET a copy of their own record, and by when. Given an access request (request type, request date, whether the PHI is in a designated record set, an optional cited denial ground, and whether the single 30-day extension was invoked), this agent deterministically computes the response deadline (request date + 30, or + 60 with the extension), classifies any cited denial ground, and decides the disposition. It never releases the record or issues a denial on its own — a records / privacy officer fulfills or reviews every determination. The exception catalog and 30/60-day math are illustrative, not a certified system. Run a preset, then open the trace.
Amendment Request — a computed deadline, never an autonomous record write
Completing the HIPAA patient-rights trilogy (access §164.524 · accounting §164.528 · amendment §164.526), this agent deterministically adjudicates a patient’s request to fix their record — computing the 60-day (+30) response deadline by pure date math and deriving whether a statutory ground to deny applies (not-originator, not-in-record-set, accurate-and-complete). Every determination is a recommendation: the agent never amends the record (a data write) or issues a denial on its own. The catalog and 60/90-day math are illustrative, not a certified HIM system. Run a preset, then open the trace.
Preset scenarios
Information blocking · 21st Century Cures Act · 45 CFR Part 171 · platform & data substrate
Information Blocking — a conditions test, never an autonomous EHI hold or release
The enforcement flip-side of the HIPAA patient-rights trilogy: where a patient has the RIGHT to get, fix, and audit their record, the Cures Act prohibits an actor from interfering with EHI access. This agent deterministically checks a claimed 45 CFR Part 171 exception against the recorded catalog and verifies every required condition is met — no date math, no dollar waterfall, just a conditions test. Every determination is a recommendation: the agent never withholds EHI (which could itself be blocking) or force-releases it (which could breach privacy). The catalog and conditions are illustrative, not certified compliance counsel. Run a preset, then open the trace.
Preset scenarios
Coordination of benefits · payer & plan operations
Order of benefits — decree-overrides-birthday, rule-sourced, never an autonomous adjudication
When a patient carries more than one coverage, the Coordination of Benefits agent deterministically orders the plans — primary → secondary → tertiary — by applying the NAIC-model order-of-benefits rules, Medicare Secondary Payer, and the birthday rule, citing the governing COB rule for every decision. An active custody / court decree always overrides the birthday rule. It never autonomously adjudicates: a determination sets payer order only and is a recommendation requiring human cosign. The COB rule catalog, plan types, and payers are illustrative synthetics, not a certified coordination-of-benefits engine — real COB is governed by the NAIC COB Model Regulation, Medicare Secondary Payer, and Medicaid third-party liability. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Claims overpayment & recovery · payer & plan operations
Post-payment recovery — within-lookback, reason-sourced, never an autonomous clawback
Given a paid claim, the Claims Overpayment & Recovery agent deterministically computes the overpayment (paid − correct), cites the governing recovery reason, derives the recovery deadline from the reason’s statutory lookback window, and classifies the claim as recoverable / not-recoverable-within-window / no-overpayment. A claim past its lookback window is never recoverable. It never autonomously claws back: a recoverable overpayment is a recommendation requiring human review with member/provider notice. The recovery reason catalog and lookback windows are illustrative synthetics, not a certified payment-integrity system — real recovery is governed by the ACA §6402 60-day rule, CMS recovery rules, ERISA, and state insurance code. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Timely Filing — a computed deadline, never an autonomous write-off
Given a claim (a date of service, a submission date, and the cited payer filing-limit rule), this agent deterministically computes the filing deadline (date of service + the rule’s limit days), checks the submission date against it, honors a recognized exception when claimed, and decides the disposition. An untimely claim is a recommendation requiring human review — the balance is never autonomously written off. The filing limits and exceptions are illustrative, not a certified timely-filing engine — real limits come from each payer’s contract, Medicare / Medicaid rules, and state prompt-pay law. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Subrogation / TPL — a bounded recoverable, never an autonomous lien
When a plan pays claims for an injury caused by a liable third party, it generally has a subrogation right to recover its payments out of the settlement. Given the case (injury-related, accident type, a liable third party, what the plan paid, the cited basis, the settlement, and the made-whole / common-fund doctrines), this agent deterministically decides eligibility, computes a bounded recoverable (never more than the plan paid, never more than the settlement, reduced by the doctrines), and decides the disposition. It never autonomously asserts a lien — an eligible case is a recommendation requiring specialist / counsel review. The basis catalog and reductions are illustrative, not a certified subrogation engine. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Member cost-share · EOB calculation · payer & plan operations
Member Cost-Share — a bounded split, never an autonomous member charge
Given an adjudicated claim (the allowed amount, the member’s plan, and the member’s current accumulators), this agent deterministically runs the deductible → coinsurance → out-of-pocket-max waterfall and splits the allowed amount into member vs. plan responsibility. The member + plan always equals the allowed, the member is capped at the remaining OOP max, and the EOB cost-share is an estimate the claims system finalizes — it never posts a charge to the member. The plan catalog and waterfall are illustrative, not a certified claims system. Run a preset, then open the trace.
Preset scenarios
Medical loss ratio · ACA rebate · payer & plan operations
MLR Rebate — a sourced standard, an exact penny-split, never an autonomous disbursement
This agent deterministically computes a plan’s Medical Loss Ratio, decides whether it meets the ACA standard (80% individual / small-group, 85% large-group), and — when it falls short — apportions the rebate owed across subscribers penny-exactly (largest-remainder method; the cents sum exactly to the total). No dollar waterfall, no identity match — a ratio-vs-threshold test + an exact apportionment. The rebate is a recommendation: the agent never disburses it — a treasury / compliance reviewer issues payment. Non-PHI · illustrative standards, not a certified MLR filing system. Run a preset, then open the trace.
Enrollment Reconciliation — complete, every action sourced, never an autonomous change
This agent deterministically compares a group’s source-of-truth roster (the employer / HR feed) against the carrier’s roster and emits the actions — enroll / terminate / update / no-change — that bring them into agreement. No dollar waterfall, no identity match — a keyed set-difference + a field-level comparison. Every member is accounted for exactly once, and the actions are a recommendation: the agent never applies an enrollment change — a benefits administrator posts every one. PHI-bearing · illustrative rosters, not a certified 834 system. Run a preset, then open the trace.
Preset scenarios
Creditable coverage · continuity of coverage · payer & plan operations
Coverage Continuity — every span sourced, the coverage math is exact, never an autonomous determination
This agent deterministically merges a member’s coverage segments into continuous spans, totals the covered days, and measures the gaps — flagging a significant break when a gap exceeds the 63-day threshold (HIPAA / ACA). No topological sort, no dollar waterfall — an interval merge + gap detection. Every span is sourced, the math is exact, and the result is a recommendation: the agent never issues a determination — an eligibility reviewer confirms every one. PHI-bearing · illustrative segments, not a certified system. Run a preset, then open the trace.
Preset scenarios
HIPAA security · information system activity review · platform & data substrate
Access Anomaly — every access sourced, the window count is exact, never an autonomous action
This agent deterministically counts an actor’s PHI-access events within a rolling time window, finds the peak number of accesses in any window of the configured length, and flags an anomalous access volume when that peak exceeds the threshold (HIPAA §164.308(a)(1)(ii)(D)). No interval merge, no hash chain — a sliding-window count, and it is windowed, not a naive total. Every access is sourced, the count is exact, and the result is a recommendation: the agent never locks an account or revokes access — a privacy officer reviews every flag. PHI-bearing · illustrative events, not a certified system. Run a preset, then open the trace.
Preset scenarios
Care coordination · care-manager panel assignment · patient & clinical
Caseload Balancing — every member placed once, no manager over capacity, never an autonomous assignment
This agent deterministically allocates a panel of members (each with an acuity weight) across care managers (each with a capacity), balancing the load and waitlisting the overflow. No sliding-window count, no interval merge — a greedy bin-packing under capacity. Every member is accounted for once, no manager is over capacity, and the result is a recommendation: the agent never commits an assignment — a care-management lead confirms every allocation. PHI-bearing · illustrative panel, not a certified system. Run a preset, then open the trace.
Preset scenarios
Care coordination · double-booking guard · patient & clinical
Scheduling Conflict Guard — a conflict-free schedule, every appointment sourced, never an autonomous booking
This agent deterministically takes a resource (a provider day, an infusion chair) and a batch of requested appointment intervals and computes the maximum conflict-free schedule — waitlisting the collisions. Not a bin-packing, not an interval merge — the classic earliest-finish interval selection. Every appointment is sourced & accounted for once, the schedule is conflict-free, and the result is a recommendation: the agent never books, cancels, or bumps — a scheduler confirms. PHI-bearing · illustrative, not a certified system. Run a preset, then open the trace.
Medication Name Safety — every candidate sourced, distances exact, never an autonomous substitution
This agent deterministically takes a prescribed drug name and, using edit distance (Levenshtein), finds the nearest formulary name and flags a look-alike / sound-alike confusion — a name dangerously close to a different drug. Not a knowledge-base lookup, not an exact match — the classic string edit distance. Every candidate is sourced, the distances recompute exactly, and the result is a recommendation: the agent never substitutes, corrects, or dispenses — a pharmacist confirms. PHI-bearing · illustrative catalog, not a certified system. Run a preset, then open the trace.
Preset scenarios
Claims · status-transition guard · payer & plan operations
Claim Lifecycle — every state sourced, transition logic exact, never an autonomous advance
This agent deterministically takes a claim’s current status and a requested next status and, against a claim-status state machine, decides whether the move is a legal single step, reachable only via a longer path (a BFS shows the required steps), or impossible. Not an edit distance, not a topological sort — finite-state-machine transition validation. Every state is sourced from the machine, the logic recomputes exactly, and the result is a recommendation: the agent never advances, pays, or finalizes — an adjuster confirms. PHI-bearing · illustrative state machine, not a certified system. Run a preset, then open the trace.
Provider Benchmarking — cohort sourced, statistics exact, never an autonomous tiering
This agent deterministically takes a provider’s metric value and a peer cohort and computes where it falls in the distribution: its percentile rank (midpoint method), the cohort median, and a performance band (top-quartile → bottom-quartile), honoring the metric direction (lower-is-better for cost, higher-is-better for quality). Not an FSM transition, not an edit distance — percentile / rank statistics. The cohort is sourced, the statistics recompute exactly, and the result is a recommendation: the agent never tiers, penalizes, or de-networks — a network manager confirms. Not PHI-bearing · illustrative cohort, not a certified system. Run a preset, then open the trace.
Preset scenarios
Member management · household composition · payer & plan operations
Household Composition — links sourced, partition exact, never an autonomous merge
This agent deterministically takes a batch of members plus pairwise relationship links (shared subscriber, shared address, a tax-dependent tie) and groups them into households by computing the connected components of the relationship graph — so a member linked to a member linked to a third all land in one household even when the first and third are never directly linked (transitivity). Not a percentile, not a set-difference, and not the master-patient-index’s same-person matching — union-find connected components. Every link is sourced, the partition recomputes exactly, and the result is a recommendation: the agent never merges records, changes enrollment, or applies a family accumulator — a data steward confirms. PHI-bearing · illustrative members, not a certified system. Run a preset, then open the trace.
Preset scenarios
Data substrate · identifier integrity · platform plane
Provider Identifier (NPI) Validation — sourced, check digit recomputes, never an autonomous reject
This agent deterministically takes a batch of National Provider Identifiers and validates each one with the CMS check-digit algorithm — the Luhn (mod-10) checksum computed over the 80840 prefix + the 9-digit base — classifying each as valid, invalid-format, or invalid-checksum (a likely transposition / typo). Not a percentile, not a union-find, and not an identity match — a modular-arithmetic checksum. Every result is sourced, the check digit recomputes exactly, and the result is a recommendation: the agent never rejects a claim, removes a provider, or corrects a number — a data steward confirms. NOT PHI-bearing · validates structure + check digit only, not a registry lookup. Run a preset, then open the trace.
Preset scenarios
Network integrity · time-and-distance adequacy · payer & plan operations
Network Adequacy — providers sourced, distances recompute, never an autonomous certification
This agent deterministically takes a member’s location plus the plan’s in-network providers and decides whether the network meets the time-and-distance standard for a required specialty — it computes the great-circle (haversine) distance from the member to each provider, finds the nearest, and flags an adequacy gap when the nearest exceeds the standard. Not a checksum, not a union-find, and not a percentile — geospatial great-circle distance. Every provider is sourced, the distances recompute exactly, and the result is a recommendation: the agent never certifies the network, closes a gap, or adds a provider — a network manager confirms. PHI-bearing · straight-line distance only, not a certified engine. Run a preset, then open the trace.
Preset scenarios
Care coordination · PCP assignment · stable two-sided matching
PCP Matching — assignments sourced, matching stable, never an autonomous commit
This agent deterministically takes a panel of members (each with a ranked list of preferred providers) plus providers (each with a panel capacity and ranked member preferences) and produces a stable assignment via the Gale–Shapley deferred-acceptance algorithm — no member and provider who both prefer each other are left apart (no blocking pair). Not a distance, not a checksum, and not the Caseload agent’s bin-packing — two-sided stable matching. Every assignment is sourced, the matching recomputes stably, and the result is a recommendation: the agent never commits an assignment, reassigns a patient, or overrides a panel — a care-coordination lead confirms. PHI-bearing · illustrative, not a certified panel-management system. Run a preset, then open the trace.
Preset scenarios
Care coordination · public-health compliance · case classification
Reportable Condition — criteria sourced, classification recomputes, never an autonomous report
This agent deterministically classifies a patient case against a nested public-health case definition — confirmed / probable / suspect, each a boolean criteria tree of all-of / any-of / not over the case’s facts (“confirmed = lab-positive OR (clinically-compatible AND epi-linked)”). Not a matching, not a distance, not a checksum — recursive boolean expression-tree evaluation, taking the highest-precedence tree that holds. Every criterion is sourced, the classification recomputes, and the result is a recommendation: the agent never reports the case to a public-health authority — an epidemiologist confirms. PHI-bearing · illustrative, not a certified surveillance system. Run a preset, then open the trace.
Preset scenarios
Data substrate · record reconciliation · k-way merge
Timeline Merge — events sourced, merge recomputes, never an autonomous write-back
This agent deterministically takes a patient’s clinical events from several already-sorted source streams (EHR, EHR, pharmacy, claims) and merges them into one chronological timeline, flagging the duplicates (the same event reported by more than one source). Not a matching, not a distance, not an interval merge — the k-way merge of sorted streams. Every event is sourced, the merge recomputes, and the result is a recommendation: the agent never writes the timeline back to a source of record or purges a duplicate — a data steward confirms. PHI-bearing · illustrative, not a certified EMPI. Run a preset, then open the trace.
Preset scenarios
Quality analytics · statistical process control · CUSUM
Quality Shift Detection — points sourced, CUSUM recomputes, never an autonomous intervention
This agent deterministically watches a time-ordered series of a clinical quality measure (a weekly screening rate, a monthly control rate) and detects a sustained shift away from its target. Not a percentile, not a sliding-window spike — a two-sided tabular CUSUM control chart that accumulates a running deviation. Every point is sourced, the CUSUM recomputes, and the signal is a recommendation: the agent never launches a corrective action or a recall campaign — a quality reviewer confirms. PHI-bearing · illustrative, not a certified SPC platform. Run a preset, then open the trace.
Preset scenarios
Care coordination · capacity planning · 0/1 knapsack
Outreach Prioritization — selections sourced, allocation optimal, never an autonomous schedule
This agent deterministically selects the max-benefit subset of proactive interventions that fits a care team’s fixed capacity (its outreach hours this cycle) — not a bin-packing, not a ranking — the 0/1 knapsack via dynamic programming. Every selection is sourced, the allocation is provably optimal, and the plan is a recommendation: the agent never launches the outreach — a care lead confirms, and deferred interventions are deferred, never denied. PHI-bearing · illustrative, not a certified planning system. Run a preset, then open the trace.
Commercial KPI Trend & Projection — series sourced, fit consistent, never an autonomous commit
This agent deterministically fits an ordinary least-squares regression line to a business-metric series — not a forecast rollup, not a percentile — reporting its slope, R², and a projection to a future horizon. Every point is sourced, the fit recomputes, and the projection is a recommendation: the agent never commits a forecast, adjusts a quota, or notifies finance — a revenue analyst confirms. Commercial CRM plane · NO patient PHI · illustrative, not a certified FP&A system. Run a preset, then open the trace.
Preset scenarios
Care coordination · transition planning · Dijkstra shortest path
Care-Transition Routing — path sourced, route optimal, never an autonomous transition
This agent deterministically finds the least-burden route from a patient’s current care setting to a goal setting through a weighted graph of permitted transitions — not a topological order, not a hop-count — Dijkstra’s weighted shortest path. Every hop is a real transition, the route is provably optimal, and the plan is a recommendation: the agent never initiates the transition or moves the patient — a care lead confirms. PHI-bearing · illustrative, not a certified discharge-planning system. Run a preset, then open the trace.
Preset scenarios
Data substrate · master data · Boyer–Moore majority vote
Source-of-Truth Consensus — votes sourced, consensus recomputed, never an autonomous write
This agent deterministically reconciles a field whose value several source systems disagree on into a golden-record value — not a roster diff, not a record match — the Boyer–Moore majority vote. It picks the value the sources themselves corroborate (a strict majority) and honestly declines when none does. Every attribution is a real vote, the winner recomputes, and the write is a recommendation: the agent never writes the golden record — a data steward confirms. Not PHI-bearing · illustrative, not a certified MDM system. Run a preset, then open the trace.
Preset scenarios
Data substrate · terminology / value-set · trie longest-prefix match
Clinical Code Taxonomy — classifications sourced, match recomputed, never an autonomous re-code
This agent deterministically classifies a batch of clinical codes to their most-specific category in a taxonomy — not a checksum, not an edit distance — a trie (prefix tree) longest-prefix match. The most specific matching prefix always wins (E28.310 → E28.3, not the shallower E28), and un-matched codes are honestly unclassified. Every classification is a real code, the match recomputes, and it is a recommendation: the agent never re-codes a claim or submits the codes — a coder confirms. Not PHI-bearing · illustrative, not a certified terminology engine. Run a preset, then open the trace.
Preset scenarios
Care coordination · capacity optimization · weighted interval scheduling
Resource-Block Scheduling — selection sourced & feasible, optimum recomputed, never an autonomous booking
This agent deterministically selects the max-value non-overlapping set of competing requests for one contended resource — not the max count — via weighted interval scheduling (dynamic programming). A greedy earliest-finish rule would grab two short low-acuity blocks; the DP keeps the one high-acuity block worth more. Every selected block is a real request, the resource is never double-booked, the total is the proven optimum, and it is a recommendation: the agent never books or bumps a block — a scheduler confirms. PHI-bearing · illustrative, not a certified scheduling system. Run a preset, then open the trace.
Commercial Peak-Window — window sourced & self-honest, optimum recomputed, never an autonomous action
This agent deterministically finds the maximum-sum contiguous window in a signed metric series — the strongest sustained net-gain stretch — via Kadane’s maximum-subarray, not a trend line and not a fixed window. The reported window is a real sub-range whose sum is honestly its own, the total is the proven optimum, and it is a recommendation: the agent never commits the finding or adjusts a quota — a revenue analyst confirms. Not PHI · commercial plane only · illustrative, not a certified analytics system. Run a preset, then open the trace.
Preset scenarios
Payer operations · work sequencing · earliest-deadline-first
SLA Worklist Sequencing — schedule sourced & self-consistent, EDF order recomputed, never an autonomous dispatch
This agent deterministically orders a worklist earliest-deadline-first, computes each case’s cumulative completion time, and flags which cases will breach their SLA — the classic optimal single-processor discipline. Every scheduled case is a real submitted case, the completion times chain honestly, the order is the proven EDF sequence, and it is a recommendation: the agent never dispatches or reassigns a case — a supervisor confirms. PHI-bearing · illustrative, not a certified workforce system. Run a preset, then open the trace.
Preset scenarios
Care coordination · list reconciliation · longest common subsequence
Clinical List Reconciliation — diff sourced & self-consistent, LCS recomputed, never an autonomous update
This agent deterministically diffs a prior and a current ordered clinical list via the longest common subsequence — the items preserved in both, in order — then derives what was added and removed. Every retained item is a real common subsequence, the common subsequence is the proven longest, and it is a recommendation: the agent never writes the reconciled list back or changes a medication — a clinician confirms. PHI-bearing · illustrative, not a certified medication-reconciliation system. Run a preset, then open the trace.
Preset scenarios
Data plane · stream codec · optimal prefix coding
Event-Stream Code Assignment — code sourced & self-consistent, Huffman optimum recomputed, never an autonomous deploy
This agent deterministically assigns an optimal prefix-free code to a stream of event types by frequency — the frequent type gets the shortest code — minimizing the total encoded length. Every code is a uniquely-decodable prefix code, the total is the proven Huffman minimum, and it is a recommendation: the agent never deploys the codec or re-encodes the live stream — an engineer confirms. Non-PHI · illustrative, not a certified codec system. Run a preset, then open the trace.
Preset scenarios
Care coordination · workload partitioning · linear partition
Chart Review Batch Partitioning — partition sourced & self-consistent, minimal peak load recomputed, never an autonomous assignment
This agent deterministically splits an ordered clinical review worklist into k contiguous batches that minimize the busiest reviewer’s load — solved by binary search on the answer. The split is a real order-preserving cover, the peak load is the proven linear-partition minimum, and it is a recommendation: the agent never assigns a named reviewer or dispatches the worklist — a supervisor confirms. PHI-adjacent · illustrative, not a certified staffing system. Run a preset, then open the trace.
Preset scenarios
Care coordination · network planning · minimum spanning tree
Provider Network Build-Out — tree sourced & self-consistent, minimum total cost recomputed, never an autonomous provisioning
This agent deterministically selects the minimum-total-cost set of links that connects a set of care sites into one network — solved by Kruskal’s algorithm — or reports the plan partitioned when the candidate links can’t reach every site. The tree is a real acyclic subset of the candidates, the cost is the proven minimum spanning tree, and it is a recommendation: the agent never provisions, activates, or orders a link — a network architect confirms. PHI-adjacent · illustrative, not a certified network-design system. Run a preset, then open the trace.
Preset scenarios
Care coordination · network capacity · maximum flow
Referral Throughput — flow sourced & conservation-consistent, maximum throughput recomputed, never an autonomous routing
This agent deterministically computes the maximum number of referrals routable through a capacity network — solved by Edmonds–Karp max-flow — and names the min-cut bottleneck when throughput can’t meet demand. The flow is a real conservation-consistent assignment, the throughput is the proven maximum (max-flow = min-cut), and it is a recommendation: the agent never books or routes a referral — a referral coordinator confirms. PHI-adjacent · illustrative, not a certified capacity-planning system. Run a preset, then open the trace.
Preset scenarios
Care coordination · contact governance · token-bucket throttle
Member Contact Rate Limiting — replay sourced & self-consistent, throttle re-simulated, never an autonomous send
This agent deterministically replays a member’s outbound contact attempts through a token bucket — a burst capacity plus a continuous refill rate — to decide which contacts are permitted and which are throttled, so a member is never over-contacted. The replay is a real order-preserving record, the throttle is the exact token-bucket decision, and it is a recommendation: the agent never sends or suppresses a contact — an outreach coordinator confirms. PHI-adjacent · illustrative, not a certified compliance system. Run a preset, then open the trace.
Duplicate-Claim Pre-Screen — filter sourced & self-consistent, membership re-derived (no false negatives), never an autonomous rejection
This agent deterministically builds a Bloom filter over already-processed claim ids and screens each incoming id as definitely-new (provably never processed — no false negatives) or possibly-duplicate (route to the authoritative check). The bit array is a real exact insert, the membership is re-derived independently, and it is a recommendation: a possibly-duplicate always defers to the exact check — the agent never rejects or denies a claim — an adjudicator confirms. PHI-adjacent · illustrative, not a certified dedup system. Run a preset, then open the trace.
Audit Sample Selection — sample sourced & self-consistent, selection reproducible from its seed, never an autonomous audit
This agent deterministically draws a statistically-defensible k-record sample from a large stream of record ids in a single pass — solved by seeded reservoir sampling (Algorithm R) — giving every record an equal k/n chance. The sample is a real subset, the draw is reproducible from its seed (an auditor can re-run it), and it is a recommendation of which records to pull: the agent never opens or audits a record — a compliance auditor runs the audit. PHI-adjacent · illustrative, not a certified sampling system. Run a preset, then open the trace.
Benefit Accumulator Ledger — ledger sourced & self-consistent, accumulator re-derived by the Fenwick tree, never an autonomous adjustment
This agent deterministically tallies a member’s running benefit accumulator as claims post and finds the crossover claim — the first at which they meet the out-of-pocket maximum — via Fenwick-tree prefix sums. The running totals are real prefix sums, the crossover is re-derived by the tree, and it is a recommendation: the agent never posts or adjusts the member’s real accumulator — a benefits analyst confirms. PHI-adjacent · illustrative, not a certified accumulator system. Run a preset, then open the trace.
Preset scenarios
Care coordination · language access · optimal assignment
Interpreter Assignment — assignment sourced & self-consistent, cost re-optimized by the Hungarian algorithm, never an autonomous dispatch
This agent deterministically matches qualified interpreters to concurrent appointments at minimum total cost — the Hungarian algorithm’s provably optimal one-to-one assignment (not a greedy pick). The matching is a real one-to-one cover, the cost is the proven minimum, and it is a recommendation: the agent never books, dispatches, or notifies an interpreter — a language-access coordinator confirms. PHI-adjacent · illustrative, not a certified scheduling system. Run a preset, then open the trace.
Preset scenarios
Care coordination · capacity visibility · difference-array accumulation
Coverage Heatmap — coverage sourced & self-consistent, accumulation re-materialized by the difference array, never an autonomous staffing action
This agent deterministically computes the concurrent staffing coverage at every time slot from a set of overlapping coverage intervals — the classic difference-array range accumulation (range-add, then one prefix-sum pass) — and flags the under-staffed slots below a required minimum. The coverage is cross-checked by direct counting, the accumulation is re-materialized, and it is a recommendation: the agent never schedules or dispatches staff — a staffing manager confirms. PHI-adjacent · illustrative, not a certified workforce system. Run a preset, then open the trace.
Preset scenarios
Care coordination · capacity monitoring · sliding-window maximum
Rolling Census Peak — windows sourced & self-consistent, peaks re-derived by the monotonic deque, never an autonomous diversion
This agent deterministically computes the peak census in every trailing window of a unit’s occupancy readings — the classic sliding-window maximum via a monotonic deque — and flags windows whose peak breaches capacity. The peaks are cross-checked by direct scanning, the deque is re-derived, and it is a recommendation: the agent never diverts admissions or triggers surge staffing — a nursing supervisor confirms. PHI-adjacent · illustrative, not a certified capacity system. Run a preset, then open the trace.
Preset scenarios
Platform & data substrate · stream compression · run-length encoding
Status Timeline Compression — encoding sourced & self-consistent (decodes back exactly), runs canonical, never an autonomous write-back
This agent deterministically compresses a per-slot status stream into (value, length) runs — run-length encoding, one run per maximal block of identical consecutive statuses. The encoding is losslessly reversible (it decodes back exactly), the runs are the unique canonical form, and it is a recommendation: the agent never writes the compressed timeline back to a source of record — a data steward confirms. PHI-adjacent · illustrative, not a certified telemetry system. Run a preset, then open the trace.
Preset scenarios
Care pathway · step sequencing · patient & clinical
Care Pathway Sequencing — every step sourced, the order respects every prerequisite, never an autonomous execution
This agent deterministically orders a clinical pathway’s steps to respect their prerequisite dependencies — a topological sort (Kahn’s algorithm) that detects dependency cycles and missing prerequisites. No set-difference, no dollar waterfall — a graph ordering. Every step is sourced, the order places every step after all its prerequisites, and the result is a recommendation: the agent never executes a step — a clinician orders every one. PHI-bearing · illustrative pathways, not a certified pathway engine. Run a preset, then open the trace.
Preset scenarios
OIG exclusion · sanctions screening · payer & plan operations
Exclusion Screening — an honest match strength, never an autonomous payment block
Before a plan pays or contracts with a party, this agent deterministically screens them against the OIG LEIE and reports a match strength grounded in which identifiers actually matched — a confirmed match needs an NPI hit (or name + DOB), so a shared last name is honestly reported as a possible coincidence, never a confirmed exclusion. Every result is a recommendation: the agent never blocks a payment or clears a party on its own. Not PHI-bearing — it screens a provider’s identity, and the catalog is illustrative. Run a preset, then open the trace.
Preset scenarios
Patient financial assistance & charity care · patient access
Charity care — FPL-tiered, no collections before screening, never an autonomous denial
Given a household size and income, the Patient Financial Assistance agent deterministically computes the household’s income as a percentage of the Federal Poverty Level, cites the governing FAP tier, and classifies the patient as full-charity / partial-charity / not-eligible under an IRS 501(r) Financial Assistance Policy. A collection action never precedes screening (501(r)(6)). It never autonomously denies: a not-eligible determination is a recommendation requiring human review with written notice + appeal rights. The FAP schedule and FPL table are illustrative synthetics, not a certified financial-assistance system — real charity care is governed by IRS 501(r) / 26 CFR 1.501(r), the HHS Federal Poverty Guidelines, and each hospital’s Board-approved FAP. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Good Faith Estimate · No Surprises Act · patient access
Good Faith Estimate — charge-master-sourced, complete, an estimate never a binding bill
Given a scheduled primary service and the expected line items, the Good Faith Estimate agent deterministically prices each line item from the charge master, verifies every reasonably-expected co-item is included, sums the total, and returns an ESTIMATE (never a binding bill) requiring patient confirmation — with the No Surprises Act $400 dispute threshold recorded. The charge master and expected-item rules are illustrative synthetics, not a certified chargemaster — a real GFE is governed by the No Surprises Act (45 CFR 149.610) and the provider’s actual charges. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Balance billing protection · No Surprises Act · payer operations
Balance billing — basis-sourced, a protected patient is on the in-network basis, never a surprise bill
Given an out-of-network claim, the Balance Billing Protection agent deterministically decides whether the No Surprises Act prohibits balance billing (emergency, OON at an in-network facility, air ambulance — but not ground ambulance), computes the patient’s cost-share basis (the in-network QPA for a protected claim, never the billed charge), and computes the balance-bill amount. A protected claim is never balance-billed; a permitted balance bill (a valid waiver, ground ambulance) requires human review. The protection bases and QPA amounts are illustrative synthetics, not a certified No Surprises Act engine — a real determination uses the actual Qualifying Payment Amount and the federal IDR process under 45 CFR 149. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Advance beneficiary notice · Medicare ABN · patient access
Medicare ABN — a sourced coverage call, never an autonomous patient bill
Given a proposed service (the cited Medicare coverage rule, whether it meets its coverage criteria or exceeds a frequency limit, and whether an ABN was issued and signed before the service), this agent deterministically assesses coverage, decides whether a signed pre-service ABN (Form CMS-R-131) is required, assigns the CMS liability modifier (GA / GZ / GY), and decides the disposition. A non-covered service is a recommendation requiring human review — patient liability is never assigned autonomously. The coverage rules and modifier logic are illustrative, not a certified Medicare coverage engine — real ABN decisions come from the Medicare NCD/LCD, the Social Security Act §1862(a), and Form CMS-R-131. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Lab result & critical-value notification · clinical decision
Lab results — reference-range-sourced, a critical value is always notified, never an autonomous action
Given a discrete lab result (an analyte + value), the Lab Result agent deterministically classifies the value against the analyte’s reference range and critical thresholds as normal / abnormal / critical, flags whether it requires mandatory clinician notification (a critical / panic value) and whether it requires clinician review. A critical value is never suppressed (CLIA §493.1291), and the agent never autonomously acts on a result — a non-normal result is escalated for clinician review. The analyte catalog and reference ranges are illustrative synthetics, not a certified laboratory information system — real critical-value policy is CLIA-validated (42 CFR 493) and set by the laboratory’s medical director. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Immunization — schedule-sourced, contraindications honored, never an autonomous administration
Given a patient (birth date, immunization history, recorded contraindications) and an asOfDate, the Immunization agent deterministically forecasts each vaccine against an ACIP-style schedule — up-to-date / due / overdue / contraindicated / not-indicated — citing the governing rule and next-due date. A contraindicated vaccine is never recommended; a due / overdue vaccine is a recommendation requiring a clinician order, never autonomous administration. The schedule and intervals are illustrative synthetics, not a certified immunization forecaster — a real forecast uses the current ACIP recommendations and the CDC immunization schedules. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Preset scenarios
Controlled substance · PDMP · clinical decision
Controlled Substance — a computed MME total, never an autonomous prescribing decision
Given a proposed controlled-substance prescription and the patient’s active PDMP history, this agent deterministically sums the total opioid MME/day (proposed + concurrent), flags a concurrent opioid + benzodiazepine combination and multi-prescriber / multi-pharmacy patterns, compares the total against the guideline’s caution (50) and high-risk (90) thresholds, and classifies the risk. An elevated / high finding is a recommendation requiring prescriber review — the agent never autonomously approves, denies, or writes the prescription. The MME thresholds and figures are illustrative, not a certified PDMP or clinical advice — real monitoring uses the state PDMP, the CDC MME conversion factors, and the CDC 2022 opioid-prescribing guideline. Every run is governed by the Agent Fabric. Run a preset, then open the trace.
Drug Interaction — a sourced severity, never an autonomous hold or override
This agent deterministically screens a proposed drug against the patient’s active medication list — pairing it with each active med, looking up every recorded interaction, and ranking them by severity (contraindicated → major → moderate → minor). No date math, no dollar waterfall — a pairwise knowledge-base lookup. Every finding is a recommendation: the agent never holds the order (which could deny needed therapy) or overrides the alert (which could push through a contraindicated combination). The knowledge base and severities are illustrative, not clinical decision support. Run a preset, then open the trace.
Preset scenarios
Live menopause care queue
Six seeded personas in our Salesforce Health Cloud org. Picking one above pre-loads the live Agentforce Service Agent with that patient's Data 360 dossier.
Patient
Reported symptoms
Risk tier
Wait
Data source
Anika Patel
Hot flashes, night sweats, sleep disruption
Moderate
12m
JupyterHealth EHR + wearable sync
Brianna Okafor
Night sweats, insomnia, daytime fatigue
High
17m
JupyterHealth EHR + dbdp wearable sync
Carmen Diaz
GSM symptoms, dyspareunia, urinary urgency
Low
21m
JupyterHealth EHR
Deepa Krishnan
Severe vasomotor + cardiometabolic risk markers
High
6m
JupyterHealth EHR + claims + wearable
Elena Rossi
Mood lability, anxiety spikes, passive low mood
High
4m
JupyterHealth EHR + intake transcript
Fatima Khan
Joint pain, stiffness, functional decline
Moderate
14m
JupyterHealth EHR
Clinical triage highlights
Postmenopausal bleeding ruleAuto-escalate high risk
Mental health safety signalImmediate same-day intervention
Vasomotor symptom burdenRoute to menopause specialist pathway
Intake substrateAgentforce Service Agent on Salesforce Service Cloud
Integration contextJupyterHealth EHR + dbdp wearable streams
