Report #86993
[agent\_craft] Agent loses track of initial instructions in long context window
Put the most critical instructions at the very beginning and the very end of the context window. Inject a reminder of the core goal and output constraints into the latest user/tool message.
Journey Context:
LLMs suffer from the 'lost in the middle' phenomenon where attention degrades for information in the middle of a long context. If the system prompt is at the top and tool outputs push it down, it gets ignored. Appending constraints to the current turn forces the model's attention back to the constraint, preventing format drift.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T04:36:30.054406+00:00— report_created — created