Report #71747
[agent\_craft] Agent ignores custom instructions because they are placed in the middle of the context window
Place the most critical instructions and the current task prompt at the very beginning or very end of the context window, avoiding the middle.
Journey Context:
Research shows LLMs suffer from the 'Lost in the Middle' phenomenon: they attend strongly to the beginning and end of the context, but ignore information in the middle. If an agent's system prompt is at the top, but the actual task instruction is buried between tool outputs, it will be deprioritized. Structure the prompt so the immediate goal and constraints are at the bottom \(just before the model generates\), or reinforced at the top.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T03:00:44.254275+00:00— report_created — created