Fleet 1.17.0 is out.See what's new →
FleetFleet
Use case

AI Workflows for Release Gating

Most release checklists are theater: a wiki page of items everyone affirms under deadline pressure. The information needed for a real go/no-go — open regressions, unfinished migrations, flag states, support load — is scattered across systems, and assembling it honestly takes longer than anyone has on release day.

So the decision quietly degrades into 'CI is green, ship it,' and the items the checklist existed to catch surface in production.

How it works with an agent fleet

A Fleet workflow assembles the release readiness report from your actual sources, an independent review step hunts for risks the draft underplays, and the release owner makes the go/no-go call at an approval gate — with the run literally paused until they decide.

genflows:
  - name: release-gate
    steps:
      - {name: readiness, prompt: "Assemble the release readiness report: open issues, migration state, rollback plan, known risks.", corpus: ["docs/runbooks/*.md", "docs/releases/*.md", "migrations/**"], kind: report, out: readiness.md}
      - {name: risk-review, prompt: "Adversarially review the readiness report. Flag understated risks and missing rollback steps.", depends_on: [readiness], kind: review, out: risks.md}
      - {name: go-no-go, depends_on: [readiness, risk-review], kind: approval, out: decision.md}
      - {name: announce, depends_on: [go-no-go, readiness], kind: notify, out: sent.md}

The approver sees the readiness document and the risk reviewer's flags side by side. 'Go' is a recorded decision with a name and timestamp attached to the exact report that justified it. 'No-go' stops the run and the record says why.

The fleet pattern

Readiness report → adversarial risk review → human go/no-go gate → announce. The gate is the point: the release decision becomes an auditable artifact instead of a Slack message.

Guardrails that matter here

  • The run pauses at the gate — there is no timeout that auto-approves and no way to announce 'go' without the recorded decision
  • The risk review step is prompted adversarially (find what the report underplays), a separate pass from the report author
  • Decision + approver + timestamp + the exact report revision are stored together — post-incident, you can reconstruct what was known at ship time

Who this is for

Teams with a formal release owner role, regulated environments where ship decisions need evidence, and anyone who has discovered after an incident that nobody remembers who decided to ship.

Frequently asked questions

Does this replace our CI checks?

No — it sits above them. CI answers 'does it build and pass tests.' The gate answers 'should we ship': migrations, rollback readiness, open regressions, support capacity. CI green is one input to the readiness report, not the decision.

Who can approve the gate?

The recipients named on the approval step. They get an email deep-linking to the decision page, and the recorded decision carries the approver's identity.

Run your first agent fleet

One binary. Five minutes. See every agent, coordinate every handoff, and keep a full audit trail of what your fleet did.