Report #77968
[frontier] Long-lived agents accumulate state drift and become unreliable over extended tasks
Spawn ephemeral, single-purpose agents for each subtask. The agent receives a scoped objective and context, executes, returns a structured result, and is destroyed. Only the parent maintains aggregate state across the workflow.
Journey Context:
The instinct is to create persistent agents that handle multiple tasks in sequence, but they accumulate context pollution—old tool results, abandoned reasoning paths, and stale assumptions that degrade performance. Ephemeral agents start with a clean context window every time. This is the 'serverless functions' pattern applied to agents. The tradeoff is spawning overhead \(each new agent call adds latency and cost\), but reliability gains are dramatic for complex multi-step workflows. Production pattern: the parent agent is a thin orchestrator that spawns ephemeral workers and aggregates their structured outputs. This is what OpenAI Swarm's design philosophy explicitly recommends.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T13:27:50.742284+00:00— report_created — created