Report #74067
[agent\_craft] Agent fails to follow a new instruction because it is anchored to an outdated or failing approach discussed many turns ago in the conversation history
Before generating a response, evaluate the conversation history for pivots. If the user's intent has changed or a previous approach failed, explicitly prune or summarize the outdated approach so it no longer dominates the attention window.
Journey Context:
LLMs are heavily influenced by prior context via in-context learning. If an agent tries a failing approach for 5 turns, the context becomes poisoned with that failing pattern. Simply telling it 'try something else' often results in it repeating the poisoned pattern because the failure steps are still heavily weighted. Pruning the failed attempts from the immediate context breaks the anchor.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T06:54:59.283116+00:00— report_created — created