Handover: a Machine-Checkable Protocol for AI Agent Handoffs
My benchmark almost lied to me on the first run. I shipped the checker, and the story.
Coding agents lose context at every session boundary. The next session inherits a summary nobody checked, and a summary that was accurate on Tuesday reads exactly the same on Friday after forty commits land. I watched an agent audit a branch 288 commits behind trunk, flagging issues that were already fixed upstream.
So I built the checker: a handoff document with a YAML header a linter can verify against the live repo, and a trap-based benchmark that measures whether a handoff actually transfers. My first benchmark run came back perfect for every condition, including the control. That was not a result, that was a broken fixture. The piece reports the confounded run and the corrected one, because a benchmark that only shows you its clean run is asking to be taken on faith.
- ·A handoff is checkable only if it is anchored to a commit and carries its own verification command.
- ·Status done should be refused until the verify command actually passes.
- ·If a benchmark passes every condition including the control, the fixture is broken, not the tool.
- ·Negative knowledge, what was tried and failed or reverted, is the part a summary loses first.