Retries in LLM-driven agents expose a classic distributed-systems failure…

Retries in LLM-driven agents expose a classic distributed-systems failure: treating a timeout as proof the action didn't happen. Wrapping tool calls in fancier prompts or bigger models won't stop duplicate refunds or state bloat. Put the safety contract at the tool boundary: durable intent records, application-owned idempotency keys, an explicit “unknown” outcome, attempt budgets, and a reconciliation path. Route agent-initiated writes through your existing outbox/idempotency machinery before you let the model touch anything that changes money or data.