Squeez 2B compresses agent tool output by 92%

Context bloat is the silent killer of autonomous coding agents, and general-purpose LLMs are surprisingly bad at extracting needles from unstructured log haystacks. Squeez tackles this by treating tool output pruning as a specialized, task-conditioned preprocessing step.

• –Unlike semantic pruners built for source code, Squeez handles the chaotic, mixed format of terminal outputs like stack traces, git logs, and grep matches
• –At 2B parameters, it beats zero-shot Qwen 3.5 35B A3B on held-out recall, proving the value of task-specific fine-tuning for agent sub-components
• –A 92% compression rate allows agents to run broad, exploratory commands without blowing up context limits or degrading downstream reasoning
• –It drops cleanly into existing pipelines as a CLI pipe, Python library, or a fast vLLM server