TRACE framework stabilizes long-horizon agents, fixing context compression instability

As AI agents tackle longer, more complex tasks, managing their context window becomes critical. Recurrent context compression—summarizing and discarding old information—is a common solution, but its behavioral impacts are poorly understood. In this empirical study, researchers (Guanghui Min, Liang Wu, Mayank Darbari, Chen Chen, and Liangjie Hong) show that naive compression weakens the influence of recent interactions, leading to increased blocked actions, repeated exploration, and instability across runs. These issues break the agent's ability to maintain coherent long-horizon behavior.

To address this, the team introduces TRACE, a verifier-guided framework that evaluates individual compaction events through paired closed-loop continuations from the same environment state. By using summary preferences, TRACE optimizes a natural-language compression prompt while keeping all underlying models frozen—no fine-tuning required. On the AppWorld benchmark, TRACE significantly improves task performance, multi-run reliability, and context-execution efficiency compared to existing compression baselines. This work provides early evidence that boundary-local evaluation—checking the impact of each compression event at the point it occurs—is a promising direction for building reliable context compression in autonomous agents.