Agentforce Service Agent
Patient-facing intake (front door) · patient-facing
Captures the structured intake record, performs red-flag screening, and produces an Open-mHealth-shaped artifact. Speaks Google A2A outbound.
Investor brief · Multi-agent control plane
Four agents, two open protocols, one governed control plane
Pause-Health.ai composes Agentforce (intake), Anthropic Claude (clinical routing), the Pause MCP server (data-plane tools), and MuleSoft (integration plane) into a single multi-agent system — orchestrated, monitored, secured, and governed by a MuleSoft Agent Fabric control plane.
The four agents
Pause Care Router (Anthropic Claude Sonnet 4.5)
Clinical-decision agent · clinical-decision
Takes the structured intake over A2A, reasons over symptoms + cycle + safety screen + age band, and returns one of six care pathways with rationale and red-flag flags. Falls back to a deterministic Pause policy engine when ANTHROPIC_API_KEY is unset or the API call fails.
Pause MCP Server
Data-plane tool surface · data-plane
Exposes the MuleSoft Experience APIs as MCP tools so any AI agent (Claude Desktop, Cursor, Agentforce Service Agent) can call get_patient_timeline, get_patient_intake, find_menopause_providers, experience_api_health as native tools.
MuleSoft Process / Experience APIs
Integration plane · integration
Three-tier API-Led Connectivity on Anypoint. The single ground-truth substrate every agent reads from and writes to. JupyterHealth + DBDP + wearables stitched into one FHIR R5 plane.
Protocols on the wire
Google Agent-to-Agent Protocol (A2A)
Agent ↔ agent handoff
Open standard from Google donated to the Linux Foundation, endorsed by Anthropic, Salesforce, MuleSoft, and OpenAI. AgentCard discovery at /.well-known/agent.json, Task lifecycle, JSON-RPC over HTTP, optional SSE streaming. Pause's Agentforce → Care Router handoff is A2A end-to-end.
Model Context Protocol (MCP)
Agent ↔ tool surface
Open standard from Anthropic now in cross-vendor adoption. Pause's MCP server (mcp/) exposes the four Experience-tier capabilities as MCP tools. The same surface is registered in Claude Desktop, Cursor, and the production Agentforce gateway.
FHIR R5 + Open mHealth
Data substrate
The clinical data crossing every agent boundary. MuleSoft Process APIs transform Open mHealth wearable payloads into FHIR R5 Observations via DataWeave; the MCP tools return FHIR Bundles; the A2A messages carry FHIR-shaped data parts.
What the Agent Fabric does
Agent registry
Every Pause agent self-registers on the fabric with its protocol (A2A / MCP / REST), endpoint, version, capabilities, governance tier, and the policies it operates under. The console at /demo/agent-fabric shows the live registry.
Policy enforcement
Twelve policies are catalogued today: model allow-list (Claude Sonnet / Opus only), no autonomous prescribing, mandatory red-flag screen, mandatory rationale, deterministic fallback on API failure, MCP tool allow-list, FHIR-R5-only substrate, mTLS for system-to-system, HIPAA audit log on every turn. Block / audit / rate-limit / redact enforcement modes.
End-to-end trace observability
Every A2A handoff and MCP tool call is recorded as a span with parent/child correlation. A patient intake span becomes the parent of the Care Router span, which becomes the parent of the MCP timeline span. The full multi-agent trace is visible in one place.
Identity-based security
Production deployments wire agent-to-agent calls through the customer's OAuth / mTLS provider via MuleSoft. Bearer tokens are issued per agent identity and validated at the Anypoint gateway before any tool call reaches the MCP server or the Care Router.
Touch the architecture
The clickable prototype runs the full multi-agent flow end-to-end. Complete an intake on /demo/intake — the Agentforce-style intake hands off to the Anthropic Care Router over Google A2A, the Care Router calls the Pause MCP server for patient context, and every span is recorded in the Agent Fabric console.
Prototype vs production
| Aspect | Prototype today | Customer deployment |
|---|---|---|
| Care Router model | Anthropic Claude Sonnet 4.5 via @anthropic-ai/sdk when ANTHROPIC_API_KEY is set; deterministic Pause policy engine otherwise. | Same SDK path, with the model selected from the customer's approved allow-list. Bring-your-own-cloud Anthropic on Bedrock / Vertex supported via env var. |
| A2A transport | JSON-RPC over HTTP (Next.js API route). No auth between agents; Agent Fabric records the trace. | JSON-RPC over HTTPS with mTLS or OAuth, brokered by the Anypoint API gateway. Identity claims propagate into the trace. |
| Agent Fabric runtime | In-memory mock (frontend/lib/agent-fabric.ts) shared across Next.js API routes. Console at /demo/agent-fabric. | MuleSoft Agent Fabric on Anypoint. Policies authored in the Agent Fabric console; trace export to Datadog / Splunk / OTel. |
| Policy authoring | Static catalog in frontend/lib/agent-fabric.ts. Read-only in the UI. | Authored by the customer's platform team in the Agent Fabric console, version-controlled, promoted across dev / staging / prod. |
| Trace store | 200-span ring buffer in-process. Survives dev-mode hot reload. | Customer's observability stack (Datadog, Splunk, OpenTelemetry). MuleSoft trace shipper exports spans with HIPAA-compliant correlation IDs. |
Phased plan
Phase 0 — Multi-agent prototype
Today
Four agents registered on the mocked Agent Fabric. End-to-end A2A handoff Agentforce → Care Router. MCP tool surface. /demo/agent-fabric console for monitoring. Live in this repo.
Phase 1 — Real Claude routing
1 week
Wire ANTHROPIC_API_KEY in Vercel (or BYO Bedrock / Vertex). Tune the system prompt with menopause clinicians. Hold the deterministic fallback in place as the safety net.
Phase 2 — First Agent Fabric customer
4–6 weeks with customer
Deploy the Care Router and MCP server behind the customer's MuleSoft Anypoint platform. Register the Agentforce Service Agent. Author the customer's policy set in the Agent Fabric console. Wire OAuth / mTLS.
Phase 3 — Multi-tenant fabric
Ongoing
Pause ships one set of agents and policies; each customer's Agent Fabric overrides what they need. Telemetry rolled up cross-customer for product analytics and clinical evaluation.
Why investors should care
- Multi-agent is the right unit of analysisPause is not 'an AI chatbot.' It is a patient-facing agent, a clinical-decision agent, a data-plane agent, and an integration plane — wired through open protocols and governed by a single control plane. The architecture matches how buyers actually operate AI in healthcare.
- Composable on open standardsGoogle A2A + Anthropic MCP + FHIR R5 + Open mHealth + DBDP + MuleSoft API-Led Connectivity. Every protocol is industry-endorsed, multi-vendor, and independently auditable. There is no Pause-proprietary glue at any tier.
- Governance is built in, not bolted onEvery agent declares its policies. Every A2A and MCP call is traced. Every decision carries provenance (which model, which path, what red-flags). This is the posture a hospital compliance officer will sign off on — not a per-agent retrofit.
- Same architecture, two product motionsB2C: patients hit Agentforce, get routed by Claude, see the right pathway. B2B: health systems install our agents on their own Anypoint + Agent Fabric and govern them. One stack, two go-to-market wedges.
Read deeper
- Agentforce intakeThe front-door agent that captures and hands off the structured intake.
- MuleSoft integrationThe connectivity plane the Agent Fabric sits on top of.
- MCP serverThe data-plane tool surface every agent calls.
- Data 360 groundingThe unified patient memory layer Pause grounds the Care Router on before every routing decision.
- Google A2A specificationThe open agent-to-agent protocol Pause speaks.
- MuleSoft Agent FabricThe Salesforce control plane Pause's deployment composes with.