Agentic Development Pipeline
One command takes a Jira ticket to a reviewed draft pull request, with humans gating every step that matters.
Most teams adopt AI tooling by handing every engineer a chat window and hoping for the best. The output is uneven, nobody can review it consistently, and the cost is invisible until the bill arrives. I wanted the opposite: a path from ticket to pull request that produced the same shape of work no matter who ran it.
The result is an eleven step pipeline. It sets up an isolated git worktree, gathers context from the ticket and the surrounding code, plans the change, implements it with tests, reviews its own diff against a written standard, and opens a draft pull request. Draft is deliberate. The ticket does not move and reviewers are not notified until a human says so.
The interesting engineering is not the prompting. It is the guardrails: where the pipeline is allowed to act on its own, where it has to stop and ask, and how you make the stopping points cheap enough that people do not route around them.
Details worth knowing
- 01Eleven step pipeline from Jira ticket to reviewed draft PR in a single command.
- 02A 28 skill reusable agent library, documented so other engineers could stand up the same workflow rather than rebuild it.
- 03Human-in-the-loop approval gates at every irreversible step, with draft-only PRs so nothing notifies reviewers prematurely.
- 04Token analytics per session and per step, so the cost of a workflow is a number rather than a guess.
- 05Isolated git worktrees per ticket, which let several agent runs proceed in parallel without fighting over the working tree.