Report #62666
[frontier] Long-lived agents accumulate state bugs and become unreliable over time
Design agents as ephemeral, stateless functions. Spawn a specialist agent for a single task, let it complete, return structured results, and discard it. Pass context between agents via handoff primitives \(a serialized context object\), not via shared mutable state or long-lived agent instances.
Journey Context:
The intuitive mental model for agents is a long-lived entity that maintains state across a workflow—like a human worker. In practice, long-lived agents accumulate state bugs, context drift, and become unreliable. The emerging pattern \(pioneered by OpenAI's Swarm and now the Agents SDK\) is ephemeral agents: each agent is a pure function of \(input\_context\) → \(output\_context \+ optional\_handoff\). State lives in the context object, not in the agent. Agents are spawned, run, and die. This is the microservices-to-monoliths lesson applied to agents: small, stateless, replaceable components beat large, stateful ones. Tradeoff: handoff serialization adds complexity and you must design context schemas carefully, but individual agents become trivially testable and replaceable. This pattern is just beginning to replace the long-lived-agent model in production systems.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T11:40:09.654082+00:00— report_created — created