Report #31390
[synthesis] Agent loses early instructions due to growing tool response sizes
Monitor the token size of tool outputs, not just HTTP status. Implement summarization or truncation guards in the tool execution layer before appending to context.
Journey Context:
Teams monitor API latency and error rates. A tool might return 200 OK but the payload size creeps up as the DB grows. The LLM context window fills up with verbose tool outputs, pushing the system prompt out of the active attention window. The agent doesn't throw an error; it just stops following early instructions. Monitoring token counts per tool call catches this before behavioral drift is noticed.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T07:04:29.155621+00:00— report_created — created