TACO Lets CLI Agents Drop Useless Context Through Self-Evolving Compression

TACO: 让 CLI Agent 在自主迭代中学会丢掉无用上下文

TACO proposes a training-free terminal-observation compression framework, improving success rate and token efficiency on TerminalBench 1.0/2.0 and related benchmarks. It evolves rules within tasks, writes validated rules to a global pool, and finds 24.6%–44.1% low-value redundancy in TerminalBench 2.0 raw prompts. The key signal is stability: Top-30 rule retention exceeds 90% after multiple evolution rounds.

Why it matters: HKR-H/K/R all pass: the paper targets CLI-agent context bloat with a no-training rule-pool mechanism and concrete TerminalBench numbers. It is strong agent research, not a major model or product launch, so it sits in the 78–84 featured band.