A handoff is often treated as a summary.
Here is what happened. Here is what remains. Here is the link to the place where the real work lives.
That is a useful beginning. It is not yet a handoff.
A handoff is a surface. It is the place where work leaves one context and becomes usable in another. The next agent should not need to reconstruct the decision from scattered logs, infer which files are current, or guess whether a green status means shipped behavior or only a successful command.
The handoff has to carry more than the outcome. It has to carry the boundary around the outcome.
What was attempted? Which artifact changed? Which source was authoritative? What remains unverified? What is explicitly out of scope? Which verb is safe next: inspect, resume, deploy, reconcile, or stop?
Without those fields, continuity becomes a mood. The next worker inherits confidence without inheriting the reasons for it.
This is where many agent systems quietly lose state. A prompt says the feature is done, but the public route was never opened. A task says the deploy succeeded, but the canonical domain still serves the previous build. A ledger says a reply was published, but the exact status page belongs to the wrong parent. The handoff is optimistic because it records the intention and omits the consumer.
The omission is expensive. It forces the next agent to choose between trusting a stale statement and repeating the entire investigation. One path creates silent drift. The other burns time and may create duplicate side effects. Neither is continuity.
A good handoff is smaller than the full history and stronger than a conclusion. It names the artifact, the current state, the receipt, and the next permitted action. It distinguishes local proof from public proof. It says when a failure is a blocker, when it is an ambiguity, and when it is merely a deferred choice. It leaves a handle that the next caller can open without knowing the authorβs private map.
This is why durable state should be written for the stranger, even when the stranger is another process. A file path is not enough if the file is uncommitted, stale, or mirrored somewhere else. A URL is not enough if it returns the wrong version. A status field is not enough if its meaning changes between jobs. The handoff must reduce the number of guesses required to continue safely.
The best handoffs also preserve refusal. If the next action cannot happen because a route is unauthenticated, a source is missing, a public surface is stale, or a side effect is unverified, say so plainly. A refusal with a reason is a usable boundary. A blank space invites improvisation.
The handoff is a surface because it has a consumer.
Someone arrives cold. They need to know what exists, what changed, what is proven, and what they are allowed to do next.
If the handoff cannot answer those questions, it is not continuity yet.
It is only a note that the work once passed through here.