Does an Agent's History Tell You When Compaction Will Hurt? A Modest, Bounded Effect on the TRACE Paired-Replay Corpus
arXiv cs.AI · 7 Oct 2026, 03:26 AEST
The study investigates whether an agent's recent behavior can predict when context compaction will negatively impact performance. Findings indicate that pre-boundary history only weakly predicts post-compaction harm.
Key facts
- Many long-horizon agents compact their context on a global rule, usually a token budget, blind to what the agent was doing.
- TRACE's public corpus of 590 harness-triggered AppWorld compaction boundaries replays each boundary from a re-executed prefix state under the pre-compaction context and under the summary, and records the burden of the next actions: calls that error or repeat a call already made.
- We find that pre-boundary history predicts post-compaction harm only weakly.
- The best frozen, interpretable trigger avoids 21% of harmful (positive-burden) boundaries while keeping 84% of compaction opportunities.
Read aloud
The study investigates whether an agent's recent behavior can predict when context compaction will negatively impact performance. Findings indicate that pre-boundary history only weakly predicts post-compaction harm. Many long-horizon agents compact their context on a global rule, usually a token budget, blind to what the agent was doing. TRACE's public corpus of five hundred ninety harness-triggered AppWorld compaction boundaries replays each boundary from a re-executed prefix state under the pre-compaction context and under the summary, and records the burden of the next actions: calls that error or repeat a call already made. We find that pre-boundary history predicts post-compaction harm only weakly. The best frozen, interpretable trigger avoids twenty one percent of harmful positive-burden boundaries while keeping eighty four percent of compaction opportunities.
Sources in this cluster
- Does an Agent's History Tell You When Compaction Will Hurt? A Modest, Bounded Effect on the TRACE Paired-Replay Corpus
arXiv cs.AI · 7 Oct 2026, 03:26 AEST