Agent Mesh: New Reliability Primitives for Agent Delegation

This research presents a failure study of a production agentic software-delivery platform, analyzing 147 incidents across 81 runs. The study found that standard service mesh reliability primitives like retry, timeout, and error-rate circuit breaking are insufficient for autonomous agents, particularly when tasks are non-idempotent. The core issue is that these primitives rely on assumptions often violated in agentic systems, leading to persistent failures, false trips, and incorrect state management. Two fundamental causes of failure were identified: "identity adequacy," where inadequate identity discrimination led to confident but wrong answers, and "evidence adequacy," where reliability decisions were made on insufficient, untraceable, or non-deterministic evidence. The study quantifies various consequences, including undetected successful tool calls, false progress signals, and misrouted failures. Based on these findings, the authors propose seven new reliability primitives, focusing on the delegation unit rather than individual messages, to improve the robustness of agentic systems.