Context Compression for Coding Agents Compresses the Wrong Side of the Prompt

A recent paper from Peking University, 'Anchored Context Distillation for Latent-Observation Software Engineering Agents,' highlights a critical flaw in how developers compress context for AI coding agents. Researchers argue that compressing an entire agent transcript—including the agent's own actions and reasoning—into latent vectors leads to behavioral drift and syntax errors. The authors propose LOHA (Latent Observations, Hard Actions), a method that keeps agent actions and recent tool outputs in raw text while compressing older environment observations. Testing on SWE-bench Verified showed that this approach significantly improves task resolution and memory efficiency compared to standard truncation or full-transcript compression. The findings suggest that developers should prioritize keeping the agent's scratchpad and tool-call history uncompressed to maintain performance, while applying lossy compression only to historical environment data.