Report #26197
[frontier] Agent context window overflows during long autonomous tasks
Implement a proactive 'context compaction' loop that summarizes older turns and compresses tool outputs before they hit the token limit, rather than truncating or failing.
Journey Context:
Naive agents append every tool output to the message history. In long-running tasks \(e.g., debugging a repository\), the context fills up, leading to truncated inputs or API errors. Simple truncation loses critical early instructions or state. The winning pattern is proactive context compaction: using a fast LLM to summarize the trajectory so far \(preserving key findings, current state, and original intent\) and replacing the history with the summary plus the most recent turns. This is triggered at ~70% context capacity, ensuring the agent never hits the hard limit.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T22:22:22.021639+00:00— report_created — created