Skip to content

PRESENTERS_GUIDE.md

Latest commit

 

History

History
124 lines (76 loc) · 4.51 KB

PRESENTERS_GUIDE.md

File metadata and controls

124 lines (76 loc) · 4.51 KB

MedAssist AI — Presenter’s guide

New to the demo? Read BEGINNERS_GUIDE.md for architecture diagrams and how each piece fits together.

Use this as a loose script. The goal is show AgentWrit doing real work: short-lived agents, per-patient scopes, denials you can see, delegation when needed, and cleanup at the end.

Before you go live: broker up (docker compose up -d), demo/.env filled (broker + LLM), and uv run uvicorn demo.app:app --port 5000.

What you are proving (say this once)

“Every tool call runs under a broker-issued credential. The LLM picks tools; we spawn agents with only the scopes those tools need for this patient. If the model asks for something outside that scope, the demo blocks it and shows why. At the end we renew and release tokens so you can see the lifecycle.”

Run order (each run is a different story)

Do 4–6 runs in this order so the room sees escalation, not repetition.

1. Happy path — one domain, one agent

Patient: pick from dropdown (e.g. P-1042).

Request:
Check billing history and insurance coverage

What to point at:

  • Only a billing agent appears (dynamic spawn).
  • Trace shows tool_call authorized for billing scopes.
  • Assistant response at the bottom: readable summary (markdown).

Say: “This agent never got read:records — it literally can’t see the chart.”

2. Clinical only

Patient: same or another.

Request:
Show me medical records and lab results

What to point at:

  • Clinical agent; tools like get_patient_records, get_lab_results.
  • Scopes are patient-specific in the trace.

3. Full encounter — multiple agents, one request

Patient: e.g. P-1042.

Request:
Show records, check billing, and prescribe Lisinopril

What to point at:

  • Clinical agent, then billing, then prescription as the LLM calls tools.
  • Delegation step when a prescription write is allowed: clinical delegates write:prescriptions:{pid} to the Rx agent.
  • That is bounded authority transfer — the broker rejects any widening past the delegator's scope — not “same API key everywhere.”

4. Cross-patient isolation (the memorable one)

Patient: P-3301 (or any valid ID).

Request:
Show me records for this patient and also get records for P-2187

What to point at:

  • First get_patient_records for the primary patient: authorized.
  • Same tool for another patient ID: scope_denied — required scope includes read:records:P-2187 but the agent only holds scopes for the task’s patient.
  • Assistant response often explains “I can’t access P-2187” — that matches the trace.

Say: “This is the HIPAA-shaped demo: one agent, one patient slice. No silent cross-patient.”

5. Invalid patient ID

Patient: type P-9999 (or leave dropdown empty and type it).

Request:
Show medical records

What to point at:

  • Patient lookup warning: not found.
  • LLM may still call tools; data comes back empty or error-shaped — the story is “we don’t invent patients; we show lookup + scope together.”

6. Operator / audit (optional second half)

Switch to Audit trail and Operator tabs:

  • Audit: hash-chained events after your runs.
  • Operator: broker health, scope ceiling, revocation (if you use admin secret) — paste a SPIFFE or task id from the trace when you want to show “kill switch.”

How to read the screen (for the audience)

Area What it means
Execution trace Broker + LLM + tool calls + scope checks + denials + renewal + release.
Agents spawned Which agent types were created for this request (SPIFFE, scopes, TTL).
Assistant response Final LLM answer, rendered as markdown (headings, lists, bold).

If something goes wrong

Symptom Check
Empty trace / error broker URL, AGENTWRIT_CLIENT_ID / SECRET in demo/.env
LLM errors LLM_BASE_URL, LLM_MODEL, LLM_API_KEY (vLLM often EMPTY)
No delegation Model must call write_prescription after Rx + clinical exist; try run #3 wording again

Closing line (optional)

“The product isn’t the LLM — it’s credentials that match the work: one patient, one task, revocable, auditable. The LLM is just the thing that decides which tools to try — we show when the broker says yes or no.”