Use Cases

Use direct agents for sprints. Use Millrace for marathons.

Direct coding agents can win sprints: small, local, well-scoped edits where a human can keep context in their head and judge the result immediately. Millrace governs marathons: work that must survive restarts, accumulate evidence, move through stages, and close under operator control.

Operator guide See the architecture GitHub

Fit 01

Use Millrace when the operating problem is bigger than the edit.

Millrace earns its complexity when work has to keep moving without relying on a single chat transcript. The runtime owns state, dispatch, evidence, recovery, and the final closure gate while agents do bounded stage work.

Good fit

Long-running

The project spans many runs, stages, or interruptions, and needs durable progress outside chat memory.

Evidence-heavy

Humans need logs, artifacts, reports, and explicit acceptance evidence before calling the root contract done.

Recovery-sensitive

The system must retry, route remediation, pause, escalate, or resume without losing the operating state.

Poor fit

Small direct edits

A single agent can inspect, patch, test, and hand back the result while the human keeps the whole task in view.

Casual use

The operator surface assumes someone is willing to inspect files, state, and evidence instead of expecting a consumer app.

Loose experiments

If failure has little cost and no closure contract, runtime governance is usually overhead.

Operational shift 02

When Millrace owns state, the operator gets a recoverable system.

Work stops being a fragile conversation. Queue records, compiled plans, active stages, artifacts, and closure evidence become the source of truth. Agents can still win sprints inside a stage, but Millrace governs marathons by deciding what runs next and whether the evidence is strong enough to close.

Run it locally Trace the handoff Inspect source