Agentic coding, ticket to merged code

Agentic coding, ticket to merged code — a self-hosted loop where I write the ticket and press merge, and two AI agents do everything in between.

I write the ticket and set it to Agent ready; a developer agent writes code and tests, checks run, a review agent reviews, and a pull request comes back to me. Blocking review findings and my /agent … comments send the work back.

What the ticket shows

The ticket moves by itself. Green: ready to merge. Purple or red: look closer. A /agent … comment sends it back to the agent.

Inside one agent run

Every "no" sends feedback back to the developer agent. After 3 rounds it stops and hands over.

Rules

rule	how
Only I start it	ticket changes and /agent comments from anyone else are ignored
Only I merge	agents work on their own branch and can never push to main
Agents can't cheat	changes to existing tests or to the CI setup are rolled back
It always stops	at most 3 rounds per run; nothing restarts without me

How this differs from CI/CD

CI/CD builds, tests and releases code that is already written. This loop writes the code from a ticket and uses CI as one of its checks. Its review agent runs before the pull request exists; the automated review in Firmware CI/CD on GitHub Actions reviews pull requests opened by anyone.

Stack

Plane (tickets) · Gitea (code, pull requests, CI) · Claude Code in headless mode (developer and review agents) · a small webhook receiver and orchestrator script.