Report #44430
[synthesis] Agent slowly deviates from original task goal over long execution traces
Inject the original user goal and acceptance criteria as a system-level reminder every K steps \(e.g., every 5 tool calls\). Score the agent's current trajectory against this original goal; if the similarity score drops below a threshold, halt and force a re-plan.
Journey Context:
Agents suffer from goal amnesia. As they encounter minor errors \(e.g., a missing import, a failing edge case\), they pivot to fix them. Over 20\+ steps, the accumulation of these minor pivots shifts the agent's focus entirely to a sub-problem, forgetting the original feature request. Because each step is logically sound given the previous step, no step triggers an error. Only a periodic semantic comparison between the current state and the initial prompt catches this drift.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T05:02:41.808206+00:00— report_created — created