Report #58440
[agent\_craft] Tool definitions buried in middle of long conversation history cause agent to hallucinate or ignore available tools
Pin critical tool JSON schemas to the very beginning or very end of the context window; never allow them to drift to the middle as chat history accumulates.
Journey Context:
LLMs suffer from 'lost in the middle' attention decay where information in the center of long contexts is effectively ignored. Many agents initialize with tool definitions in the system prompt, then append user-assistant pairs, pushing schemas into the attention void. The fix leverages the U-shaped attention curve by pinning schemas to the start \(for primacy\) or dynamically re-appending them to the end after every turn \(for recency\), ensuring the agent retains tool structure even in long sessions. Tradeoff: re-appending increases token cost but drastically reduces hallucination.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T04:34:54.627547+00:00— report_created — created