Why AI Agents Must Treat Network Timeouts as 'Outcome Unknown', Not Failure

When an AI agent's HTTP request times out, recording the result as either 'failed' or 'succeeded' creates a dangerous blind spot in automated systems. A timeout after data is sent does not confirm whether the remote server processed the request, meaning a blind retry can produce duplicate payments, tickets, or emails. A proposed state machine introduces 'outcome_unknown' as a first-class status that blocks automatic retries and triggers a reconciliation loop to verify the actual server outcome. The framework distinguishes pre-send failures, which are safe to retry, from post-send ambiguity, which requires external proof of effect before any action is repeated. For APIs lacking native idempotency key support, the approach recommends that callers generate a canonical cryptographic fingerprint of the intended action to enable safe deduplication.