Case study

An agent working our backlog.

We don't sell vapor: we run what we build. This is our own agent, working our own Jira, with real numbers.

The numbers

163

autonomous cycles in 3 weeks

32

PRs reviewed, 4 review follow-ups

13

backlog tickets closed

190s → 79s

median cycle time after optimization

The loop, live

  1. 01

    Jira backlog

    prioritized tickets

  2. 02

    Claude Code agent

    autonomous cycles

  3. 03

    PR + tests

    never straight to main

  4. 04

    Automated review

    approve if only nits

  5. 05

    Merge + evidence

    logs, diffs, decisions

The rules that made it work

01

Human-in-the-loop by design: the agent escalates whatever needs judgment; it doesn't improvise.

02

"Approve if only nits" — and "nit" is defined in writing. Without a definition, the rule is a roulette.

03

Everything goes through PRs with tests, never straight to the default branch. Autonomy lives inside guardrails.

04

Every cycle leaves evidence: logs, diffs, decisions. If it can't be audited, it doesn't ship.

The stack

  • Claude Code
  • Jira
  • GitHub

Claude Code + Jira + GitHub, orchestrated with our own evals and scheduling framework.

This is exactly what we set up for clients.

Agents with evals, guardrails and evidence. If you want one working your backlog: