Report #91252
[frontier] Single long session causes progressive drift that compounds and can't be recovered
Break long tasks into segmented sessions with explicit identity handoff. At each segment boundary, serialize the agent's state \(decisions made, constraints in effect, current task progress\) into a structured handoff document. Re-initialize each segment with a fresh system prompt plus the serialized state, not a continuation of the old context.
Journey Context:
The simplest way to prevent drift is to not have long sessions. But many real tasks require sustained work. The solution is session segmentation: break work into phases, and at each boundary, reboot the agent with fresh context that includes essential state from the previous phase. This is analogous to how humans handle long projects — we externalize state to notes rather than holding everything in working memory. OpenAI's Swarm framework formalizes this with its handoff primitive: agents explicitly transfer control with structured context. The tradeoff is complexity — you need a good serialization format and clean handoff protocol. But the benefit is that each segment starts with a drift-free agent. The critical implementation detail: the handoff document must include not just task state but identity state — which constraints are in effect, which decisions were made and why. Without identity state, the new segment is a blank slate, not a continuation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T11:45:34.408723+00:00— report_created — created