Case Study
Agent work coordinator
99 tools agents use to plan, claim and close work
Custom production app · AI operations
The goal
Let one operator run several AI agents against one backlog, with the operator's time spent on decisions rather than on routing work and repeating context.
The problem
Agents could do real work, but coordinating them fell on one person. Each new agent needed the same context repeated. Work collided, or disappeared when an agent crashed or ran out of quota. An agent's "done" was only a claim, so every result had to be checked by hand.
What we did
We built a self-hosted tracker (teams, projects, issues, statuses) whose agent interface mirrors a mainstream issue tracker's API, so an agent already wired to that tracker switches over with a config change. The web API and the agent tools share one set of handlers and can't drift apart.
Leases guarantee two agents never take the same task. Before starting, an agent pulls one execution packet with its instructions and current gates. When an agent gets stuck, it hands the work back with a reason instead of guessing. Completion is backed by recorded evidence rather than the agent's word.
The results
- ✦99 agent tools and 1,342 commits between May and September 2026
- ✦More than 2,500 automated tests (2,265 back end, 272 front end)
- ✦In a pilot, a fresh agent built a ticket correctly from the tracker's brief alone, with no person translating the task. Delivery outcomes are still being measured.
What it does for revenue
- Margins
- One operator supervises several agents because the agents route their own work and escalate only real decisions. Output grows faster than headcount.
- Speed to market
- Agents work the same backlog in parallel without colliding, since a lease lets only one agent hold a task at a time.
- Cost
- Self-hosted, with an API that existing agent tooling already speaks, so moving onto it costs a config change and leaves no vendor lock-in.