In a long trajectory, tool calls fail routinely: rate limits, stale ids, malformed arguments, empty results. An agent's competence is largely its competence at recovery, so failure paths deserve as much design attention as the happy path.

University of Michigan - Ann Arbor

A topic of the Building AI Agents course. Planned for 2 hours at Hard difficulty.

Failure Handling in the Tool Loop

Failure Is the Normal Case

An idempotent tool can be called twice with the same arguments and leave the world in the same state. Since retries are unavoidable and the agent cannot always tell whether a timed-out call landed, write-side tools should be idempotent by construction, usually via a caller-supplied key.

Idempotency

delay = min(base * 2^attempt, cap) * (0.5 + random()/2). The jitter matters more than the exponent when many agent instances retry at once, because synchronised retries reproduce exactly the overload that caused the failure.

Learn Before

Related