TACO: Self-evolving compression for terminal agents cuts token cost by 10% while boosting accuracy 1-4%

TACO, a plug-and-play compression framework for terminal agents, automatically discovers and refines compression rules from interaction trajectories, reducing token overhead by ~10% while improving performance by 1-4% on TerminalBench and generalizing across six benchmarks. The method addresses redundancy in long-horizon agentic tasks by learning task-aware compression without heuristic prompts. Code is not yet released, and gains are benchmark-specific.

• #agent
• #compression
• #terminal
• #taco