Use cases/Review

A review queue where decisions feed back.

Reviewers tag Accept · Expected · Dismiss. Two Accepts on mon_attending_approval become a guardrail candidate. Expected silences the recipe for that entity shape. Every decision is attribution + audit.

See the product
Before
Same finding triaged twice — once by morgan.lee, once by PharmD Okafor
→
With Invariance
Two Accepts auto-draft mon_attending_approval guardrail

Walkthrough

How a run becomes a guardrail.

01

Findings land with evidence

Severity, affected entities (patient, policy, escalation), cited trace nodes, recommended action — already on the card when the reviewer opens it.

02

One-click decisions

Accept · Expected · Dismiss. Each decision is attributed (which pharmacist, when, note) and tied to the same graph as the agent's run.

03

Decisions train the system

Accepts → guardrail candidates. Expecteds → recipe-silencers for OB-GYN-style nurse-station overrides. Dismisses → severity-model signal.

Built on these surfaces

The product behind this workflow.

Reviews

One queue. Attribution per decision.

Learn more →

Operators

Pharmacists, attendings, and agents in one registry.

Learn more →

Divergence

Findings flag where execution left intent.

Learn more →

Keep reading

Related use cases.

Ship
Ship safer agents
Guardrails the system writes — from your own reviewed runs.
Monitor
Catch silent failures
200 OK and the attending sign-off is missing. We name the missing edge.
Debug
Debug agents in production
From a Floor-7 escalation to the red edge in one run ID.

See it on your own workflow.

15 minutes. One workflow. No deck.

All use cases