Report #46355
[frontier] Long-lived agent instances accumulate state drift and context bloat over multi-step tasks
Spawn ephemeral agent instances for each discrete subtask, passing only the distilled context needed for that specific task, and destroy them after completion—treating agents like serverless functions, not persistent servers.
Journey Context:
The intuitive approach is to create an agent instance that persists throughout a complex task, accumulating context as it works. In production this leads to: context windows filling with irrelevant prior steps \(the agent spent 10 turns on a dead-end branch and now that history dominates the window\), agents being influenced by abandoned approaches \(they 'remember' a prior strategy and keep referencing it\), and inability to parallelize subtasks because state is coupled to a single instance. The emerging pattern is ephemeral spawning: create a fresh agent for each subtask with only the context it needs. This is analogous to serverless functions vs long-running servers. The tradeoff is overhead from repeated system prompt processing and initialization, but this is offset by smaller per-turn context windows \(cheaper and faster inference\), clean state with no drift, and the ability to run subtasks in parallel. OpenAI's Swarm framework uses this pattern implicitly—each handoff creates a fresh agent context rather than accumulating state in a single instance. The implementation key: define a clear 'task context' type that gets passed to each spawned agent, containing only what that agent needs and nothing from prior agents' internal deliberation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T08:16:52.493051+00:00— report_created — created