Agent release train

Release train of agents

Releases go wrong in the gap between "the code is merged" and "the artifact is live": an undocumented breaking change, a fresh CVE in a bumped dependency, a rebuild that was never the thing that passed. A pipeline of release agents closes that gap by making each step produce a record: what changed, what it risks, whether it verified, and only then does a human approve the publish. The conductor holds the gate; the agents fill it.

Team

• Conductor (release-manager-role): cuts the manifest and holds go/no-go.
• Changelog agent (technical-writer-role): drafts the release notes.
• Risk scanner (security-engineer-role, dependency-auditing): flags hazards.
• Verifier (qa-engineer-role): confirms the candidate is green.

Shape: a sequential pipeline ending at a judge-style go/no-go gate.

Method

1. Cut the manifest. The conductor fixes the commit range and version from the diff since the last tag into manifest.md, with the cut criteria stated up front.
2. Generate the changelog from history. The changelog agent groups merged PRs into features, fixes, and breaking changes in CHANGELOG.md, and flags every breaking change explicitly, not buried in a bullet.
3. Scan risk over the diff. The scanner lists new or bumped dependencies, known CVEs, database migrations, and config changes into a ranked risk.md, with a blast-radius note per item.
4. Verify the exact candidate. The verifier confirms required checks are green and smoke tests pass on the built artifact, never a rebuild, writing pass or fail per check to verify.md.
5. Run the go/no-go gate. The conductor reconciles the three files: any open blocker, unflagged breaking change, or failed check holds the train and routes to the owner. No selective reinterpretation at 5 p.m.
6. Publish behind a human approval. A person approves the actual tag, publish, or deploy; ship the artifact that passed, promoted gradually through canary or rings, watching error rate and latency.
7. Record the go-live. Append the go/no-go decision, attendees, and rollout result to manifest.md so the 2 a.m. on-call reads a real record.

Run it

In Claude Code, run the four agents as subagents over one shared release directory; the orchestrator sequences changelog, risk, and verify (those three can run in parallel), then performs the go/no-go itself and stops for human approval before any publish command runs. Port it to CrewAI as a sequential process with a gating task, to AutoGen as a GroupChat whose conductor approves transitions, or to LangGraph as a linear graph with a conditional edge that halts on any red signal.

Signals it works

• Breaking changes appear flagged in the changelog, not discovered by users.
• The published artifact is byte-identical to the one the verifier passed.
• A held train routes to a named owner instead of shipping around the gate.

Boundaries

This automates the release record and the gate, not the fix for what the gate catches, which goes back to the owning engineers. The cut criteria, rollout policy, and who may approve a publish are company convention the conductor follows, not invents. A human holds publish and rollback authority; agents prepare the decision, they do not execute the irreversible click.