Product fit / when to use
In developmentUse Millrace when the work needs a runtime
Direct agents suit small edits that a human can judge immediately; Millrace suits long-running work that needs durable state, evidence, recovery, and governed closure.
Page state reviewed 2026-08-15. The page does not expand the release evidence.
Good fit
Use Millrace when work is long-running, evidence-heavy, restart-sensitive, or spread across multiple bounded stages. The operating problem is bigger than a single edit: progress must survive interruption, results must be checked against a declared graph, and completion must be accepted through a legal transition.
Poor fit
Use a direct agent for a small, local change that can be inspected, patched, tested, and handed back while the human keeps the entire task in view. Casual experiments and loose prototypes usually do not need a durable workflow runtime.
The practical boundary
Millrace does not make a runner omnipotent. It governs the selected graph, evidence, and transitions while the local runner still has the permissions of the operator account that launched it. Its value is recoverable coordination and explicit authority, not a claim of operating-system isolation.
The architecture guide explains the boundary in source-backed terms.