Tutti

The problem

Running one coding agent is easy. Running several is not. The moment you have more than one, the hard part stops being the model and becomes coordination: which agent runs next, in what order, what work blocks what, and when a human needs to step in before something risky lands.

The approach

Tutti is a local coordinator that makes one deliberate choice: it does not invent a new control plane. Every unit of work is a GitHub Issue, and the state lives in labels (state:*, role:*, type:*, priority:*, override:*). The coordinator polls, reads the labels, and dispatches the right agent. A Dispatcher classifies fresh issues, then specialised roles take turns: researcher, PM, engineer, QA, and a merger that lands the PR.

Technical challenges

• Safe handoffs. Each role writes something durable and flips labels to pass work on, so the workflow is a state machine whose state I can actually see.
• Knowing when to stop. Dependency gates hold work until blockers clear, human-verification pauses wait for a person at the risky points, and retry and loop detection catch agents that get stuck spinning.
• Forcing past gates safely. Operator overrides are explicit labels, so bypassing a gate is always a deliberate, visible act.

Outcome

A working coordinator where the agents and I look at the same board. When something goes wrong I am debugging in GitHub, a tool I already know, instead of a bespoke dashboard I would otherwise have to build and maintain.

What I’d do differently

Generalise the issue backend earlier. It started GitHub-specific, and pulling the workflow logic away from the backing service is cleaner as an adapter from day one.