An AI teammate that works the ticket backlog
outcome one epic per fortnight became four epics in parallel
Problem
A data team’s throughput was capped by one engineer’s hours. Tickets queued for days, and small well-specified tasks waited behind big ones.
What I built
A three-skill workflow around the team’s real Jira board:
- My tickets looks at everything in flight and in the backlog for the developer invoking it, ranks what to work on next, and, where a ticket is only a vague sentence from a colleague, runs a full interview-and-adversarial-review cycle to turn it into a real specification before anything is built.
- Decompose breaks an approved project into roughly one-day tickets, explicitly split into work agents can do alone and work that needs the developer.
- Orchestrate points at an epic, runs the automated tickets end to end (coding, documentation, repo work, through to a pull request), and walks the developer through the manual ones.
The agent follows the team’s real ticket lifecycle: moves tickets to In Progress when starting, writes a plain-English comment for stakeholders, transitions to In Review on completion, and verifies each transition actually took effect rather than assuming. Unspecced work is skipped, not improvised.
Techniques that made it work
- Tickets deliberately decomposed to one-day units or smaller, so AI throughput is measurable on the board like any team member’s
- A companion plan-decomposition skill that turns an approved design into cohesive, independently reviewable tickets; the AI is only allowed to create work items from an examined, approved plan
- Guardrails: unspecced work is skipped, not improvised, and every status change is verified after the fact
Outcome
The bottleneck moved to where it belongs: human judgement. An agent can clear as many as 50 well-specified tickets in a day; what limits throughput is specifying and reviewing, which is exactly the work a developer should be doing. In practice a developer who used to spend one to two weeks on a single epic now runs three or four in parallel, one being specified, one being orchestrated, one being tested, and finishes them in days. And because every plan survives interviews and adversarial review before code exists, more consideration goes into each project than the old pace ever allowed.