Protocols for handing work between agents under partial context
Most handoffs lose state. The interesting part is what survives — plan, constraints, and in-flight edits carry across a handoff far more reliably than raw conversation history does.
An agent that runs out of context, quota, or patience mid-task has to hand its work to a fresh agent that knows nothing about what came before. The naive fix — dump the whole transcript — is expensive and still loses the thread.
A signed continuation contract (intent, plan, remaining tasks, in-flight code, decisions and constraints) turned out to compress the handoff to what actually matters, and it survives crossing between different model providers, not just a fresh session of the same one.
The harder part isn't the schema — it's timing. A handoff triggered after the agent has already stalled is a recovery. A handoff triggered from a forecast of remaining capacity, before the wall arrives, is a relay. The latter reads completely differently to whoever's waiting on the output.
Still open: how to represent partial trust — when the incoming agent should verify a claim in the contract rather than take it at face value.