Arcane Architecture
Doctors often track follow-ups from memory or sticky notes. If a follow-up slips, the patient may never get checked on. When they do call back, the doctor has to reconstruct the patient's condition from scratch.
Arcane fixes both problems: a short note after each visit creates a fast, searchable patient history, while follow-ups are scheduled once and notified automatically when they are due.
The Workflow
A doctor sets a follow-up once, such as "check back in 7 days." From there, everything happens automatically.
- The doctor records a short note after the visit.
- The note becomes searchable immediately.
- The doctor schedules the follow-up duration.
- A Step Functions wait holds the workflow until the follow-up is due.
- The reminder agent reads the relevant notes for context.
- Cedar authorizes the agent's tool call.
- The patient receives the follow-up reminder.
The next call starts with context instead of relying on memory.
The AWS Stack
Built on the Build It track, using AWS open-source tooling locally with no AWS account required.
The same application code maps directly to real AWS services in production, without requiring a rewrite:
- LocalStack → DynamoDB and Step Functions
- AWS SAM CLI → Lambda and API Gateway
- Cedar → authorization policies
- Strands Agents SDK → reminder agent
- OpenSearch → patient note search
- Finch → local container runtime
For production, the local components can be replaced with:
- Amazon OpenSearch Service
Cedar remains the same.
Why Cedar, Not Just a Prompt
Patient notes are sensitive. A doctor should only be able to access their own patients.
Arcane enforces that rule with a Cedar policy on every request instead of relying on application-level if checks that could be forgotten when a new endpoint is added.
The same authorization model also limits what the reminder agent can do.
The agent is built with the Strands Agents SDK, and every tool call is checked by Cedar before execution.
The agent can:
- Send a follow-up reminder
The agent cannot:
- Access another doctor's patients
- Call tools outside its allowed scope
These restrictions are tested directly without running an LLM. Calling the agent's tool functions in-process triggers the same Cedar authorization check used during a live agent conversation.
That makes the guardrails enforceable by policy, testable in isolation, and independent of the model's behavior.
Loading article