Paper: agent failures stem from infrastructure, not hallucinations

An ACM Queue paper by Tommy Tran, "Failure Is a Feature," argues that most AI agent failures are distributed-systems problems — timeouts, truncated payloads, lost acknowledgments and loops — not model hallucinations. In the paper's test harness, a bare agent completed 33.3% of tasks, retries alone reached 83.3% but created duplicates, and all seven reliability patterns hit 89.6% with no duplicates.

• Bare agent completed 33.3% of tasks; retries alone hit 83.3% but created duplicates
• All seven reliability patterns reached 89.6% with no duplicates or orphaned writes
• Idempotency keys should be derived by the harness, not filled in by the model
• Step, token and time caps are called the cheapest insurance