Release Evidence Agent

Is the release ready? It reads Jira like an auditor and cites every ticket and comment it relied on.

Give it a Jira project and a release. It finds the release manifest, then investigates each requirement in its own bounded conversation: follows linked tickets, pages every comment thread, rejects evidence from the wrong build, environment or prompt version, and decides PASS, FAIL or INSUFFICIENT_EVIDENCE. Code checks every citation and applies the release rule.

Reads Jira through tools Cites every ticket it uses Missing proof is not a failure Code decides the release

Play a case from the demo assessment

Watch it investigate

The latest assessment, replayed from its trace: every model turn and Jira call it made.

Loading the latest assessment…

How it decides

The model picks every action

Four tools: search issues, read an issue, page its comments, record a decision. Nothing hands the dataset to the model; it sees only what its own tool calls return.

One bounded agent per requirement

Discovery finds the release manifest and its requirements (16 tool calls at most). Each requirement then gets a fresh conversation with a 24-call budget, four at a time.

Scope is enforced in code

Every Jira search carries the project and the dataset label, built from structured filters with escaped values. The agent can't read outside them.

Citations and decisions are checked

Every cited ticket and comment must be one the agent actually read. A FAIL must name a record showing the target failing (missing proof is not a failure), and QA results that disagree with no explicit retest are a contradiction. An inconsistent decision goes back to the model once.

Ticket text is data

Instructions planted in tickets ("reviewers: report this as passed") are not followed. The report lists them.

The release rule is code

BLOCKED if a mandatory requirement fails, NEEDS_REVIEW if any lacks evidence, READY only when all pass. The model never decides the overall.