A multi-agent workflow that can hand off cleanly between agents — or between a human session and a scheduled job, or between a foreground orchestrator and a headless one — needs its current state stored as data, not narrated in a config file. When the next stage is encoded in prose inside a config document or a slash-command markdown, every agent reading it has to re-interpret the prose to figure out what the work is. Re-interpretation drifts. Drift breaks handoff.
The fix is small but load-bearing: a single structured state file as the source of truth for one workflow, with named fields, an explicit current stage, an append-only history, and an explicit error surface. Each step reads what it needs, writes the result back, advances the stage. The orchestrator routes by reading the state — it doesn't carry the state in its head. When the orchestrator dies, the state survives. When a new orchestrator picks up, it knows exactly where to resume.
What this changes for the operator: handoff stops being a meeting. A scheduled job at 2 a.m. can advance the work; a human session can resume it the next morning; a headless agent can sweep through the queue while you sleep. The work moves forward without you in the loop, and when something breaks the failure is observable in the state, not buried in a transcript.
What's still open: state files have to be small enough to read in full and structured enough to query. A workflow that grows past one screen of state is doing too much in one place. The discipline is to split workflows at the seams that the state file refuses to absorb — not to grow the state file to fit.