Answers
Handing work between your own AI routines, without re-narration
Scheduled AI routines usually pass work to each other by retelling it — a note in a file, a message in a channel, last night's transcript — and a retold fact drifts. With Arroway a run that stops unfinished leaves an explicit handoff: where it stopped, the one next step, the risk left open, and who it is for. That handoff is served first to whatever reads next, and it stays there until someone closes it. Nothing is inherited as a summary of a summary, and nothing quietly falls off the list.
Last updated September 3, 2026
A retold fact changes shape, and nobody sees the hop
The first run says a migration is pending. The second reads that and writes that the migration is blocked. The third reads that and asks who to escalate to. No step was dishonest, no step went back to the original, and by the third telling the team is solving a problem that does not exist. This is the ordinary failure of passing work by narrative: every hop is plausible, the origin is three hops back, and nothing marks the distance.
A handoff has an owner and an ending
When a run stops without finishing, it writes down four things: where it stopped and what is already verified, the one concrete next step, the risk left open, and who the handoff is addressed to. It is served first to whatever reads that project next, and it does not expire by clock — it ends when someone closes it, replaces it on purpose, or discards it. A run that finishes writes something different: what the work left behind, so nobody arriving later redoes it.
What is finished and what is in flight are different records
Most of what a run produces is a result: this was decided, this was measured, this changed. That belongs in a dated record, and it ages naturally. In-flight work is the exception, and it has to behave differently — it has to keep appearing, and it has to be closeable. Mixing the two is how items disappear: a half-finished task written as a result reads like a result, and nobody comes back for it. And a run that writes neither is indistinguishable from a run that never happened.
What it looks like in practice
A nightly routine gets through four of ten items and stops. It leaves a note in a file it wrote for itself. The next night's run starts fresh, reads nothing, and redoes the first four. Six items are never done, and nobody knows they exist. With a handoff, the same night ends differently: four items done and verified, item five is the next step, the open risk is that item seven depends on a decision nobody has made, addressed to whoever picks it up. The next run opens with that on top. It starts at item five, and item seven becomes a question for a person instead of a silence.
Questions people ask about this
- Why not just have the next run read the previous transcript?
- Because a transcript is the whole session, and what the next run needs is the few lines that matter. Reading everything means re-deriving the conclusion, and re-deriving is where it changes. A handoff is written on purpose by the run that stopped, which is the only moment anyone knows what was verified and what was merely assumed. An hour later, from the outside, that difference is not recoverable.
- Doesn't an old handoff just sit there forever?
- It sits there until someone deals with it, and that is deliberate. Age is shown, not acted on: a handoff from three weeks ago arrives labelled as three weeks old, so you can judge it, and it is never removed because time passed. It ends in exactly three ways — closed as done, replaced by a later handoff that names it, or discarded by a person. Anything that expired on its own would take the unfinished work with it, silently, which is the failure this exists to stop.
- Do my routines have to be written a particular way for this?
- They have to open by reading and close by writing, and that is the whole requirement. A routine reads what applies when it starts, and ends one of three ways: a record of what it settled, a handoff if it stopped early, or an explicit line saying nothing durable came out of it. That last one matters as much as the others — a run that says nothing looks exactly like a run that did not happen, and the next one has no way to tell which it was.
Where this is verifiable
Product documentation on this site (How it works, Install), the sanctioned product spec in the repository, and the Privacy Policy for what is stored and who can reach it. Everything described here is behaviour the tools apply today, not roadmap. Nobody has to remember to fetch a handoff or open another tool for it; how firmly a routine's opening read can be required depends on the AI tool it runs in.
https://www.arroway.app/en/answers/handoff-between-your-own-scheduled-ai-routines