When a Tool Times Out: Recovery Contracts for Stateful AI Agents

A tool timeout leaves an AI agent with two uncertainties: whether an action happened and whether repeating it is safe. A language model can generate a plausible recovery plan while creating a second purchase, refund, message, or infrastructure change. This white paper proposes a recovery contract that separates model planning from durable action identity, authoritative outcome reconciliation, and permission to resume. It specifies a state machine, adapter requirements, and a fault-injection evaluation protocol focused on externally visible side effects. The proposal combines established distributed-systems principles with recent agent-reliability evidence; it does not claim a new transaction primitive. No implementation or empirical evaluation is reported. Its practical contribution is a testable architecture for making recovery decisions auditable and for distinguishing safe suspension from successful completion.

Artificial Intelligence Cloud Computing Software Engineering