Summary
Software engineering agents often work by trying many things, which creates very long records of their steps that are hard to manage and expensive to process. The researchers developed AttnCompress, a method that smartly shrinks these records by focusing on the most important parts without losing key details. Their approach uses signals from the agents’ attention and patterns in the code to decide what to keep and when to revisit past information. Tests showed AttnCompress keeps important information better than older methods, cuts processing needs, and works well with different programming languages. This helps software agents handle complex tasks more efficiently.
software engineering agentstrajectory compressionattention mechanismcontext windowperplexitydynamic rolling windowsemantic relevancetoken consumptionprogramming languages
Authors
Zhengran Zeng, Yixin Li, Rui Xie, Wei Ye, Shikun Zhang
Abstract
The transition from human-centric assistance to Autonomous Software Engineering (ASE) agents has enabled the resolution of complex real-world SE tasks. However, the trial-and-error nature of these agents generates lengthy interaction trajectories, creating severe bottlenecks in terms of context window limits and cost. While context compression offers a potential remedy, prior approaches suffer from static pruning strategies and granularity mismatches, often failing to preserve the semantic dependencies and syntactic details crucial for SE tasks. To strictly preserve critical task evidence while reducing context length, we introduce AttnCompress, a dynamic attention-guided trajectory compression framework. Unlike existing approaches, AttnCompress bridges the gap between semantic integrity and dynamic adaptability through three key mechanisms: (1) structure-aware segmentation via perplexity (PPL) spikes to preserve the syntactic structure of code and logs; (2) relevance estimation using proxy attention weights to quantify the precise relevance of historical blocks to the agent's current reasoning; and (3) a dynamic rolling window to re-evaluate and recall historical context as the task evolves. Extensive evaluation on SWE-Bench-Verified and Multi-SWE-Bench demonstrates that AttnCompress achieves a pass rate of 53.17%, outperforming prior state-of-the-art baselines while reducing token consumption by 21.6% and total costs by 33.6%. The framework proves to be model-agnostic and generalizes effectively across diverse programming languages.