Flow is an opinionated durable workflow for consequential multi-step changes, in preview, for people who will read the review.
That sentence is the product boundary and the honest limit at the same time. Flow adds ceremony — a plan you approve, one feature at a time, host-observed validation, an independent review, an explicit close — and that ceremony is worth paying for only when a wrong change is expensive. For anything else it is overhead.
- The change spans several files or steps and has an order that matters.
- Being wrong is costly: a migration, a lifecycle invariant, a security boundary, a release path.
- You want durable state that survives a restart, a compaction, or a lost transcript, and a record of what was actually observed.
- You will read the plan before approving it and the findings before accepting a pass.
- The change is small. A one-file fix, a rename, a typo. Flow will plan it, validate it, review it, and close it, and every one of those steps costs a model turn you did not need.
- You are exploring. Flow locks a plan on approval and refuses to fold a materially different request into an active goal. That is the wrong shape for "let's see what happens if".
- You want speed above all. A serial lifecycle with an independent review is slower than asking directly, by design.
- You will not read the review. Flow's review is a real model judgment, not a proof. Treating a passing verdict as a guarantee is the one way to be worse off than not using Flow at all — see guarantees for exactly which parts the runtime enforces and which parts are judgment.
- You need parallel features, multiple repositories, or team orchestration. Flow runs one durable feature at a time in one project. Parallelism is bounded to workers inside a single feature.
- Your host cannot carry
/flow-autocontinuation. It still works, one/flow-runat a time; Flow tells you at startup when it detects this.
Flow is pre-1.0 in the sense that matters: its reliability is measured, but not yet measured enough to promise. What that means concretely:
- Evals run against multiple providers on a schedule, and the numbers are published with the release rather than described. Absent numbers mean unmeasured.
- Session v5 is the only active state contract. There is no migration or rollback layer: finish or close active work before changing versions, in either direction.
- The public surface is frozen while the guarantees are measured. See the release policy in release qualification.
- External validation is thin. Flow has been used seriously by its author and little else, so "it works" currently means "it worked here".
Claims about Flow should be traceable to a test, a scheduled eval, or nothing. If a claim is in the last category, this document is the place it gets said out loud.