Distil, a Context Compression Tool with No Agent Performance Degradation, Released

Distil, a context compression tool that introduces statistical non-inferiority testing (TOST/McNemar) to prevent decision changes that can occur when compressing an agent's context, has been released.

Key features are as follows:

• Decision-equivalence guarantee: Rather than simple byte compression, it performs compression while verifying that the agent's tool calls and output results remain unchanged.
• SWE-bench validation: Testing on 500 SWE-bench Verified tasks demonstrated that the performance difference between using full context (39.2%) and using compression (42.0%) was not statistically significant.
• Cache efficiency optimization: The front part of the conversation (Prefix) is kept stable at the byte level to improve API prompt caching efficiency, while compression is performed mainly on the back part (Tail).
• Versatility: It runs on Python 3.9+ with no separate runtime dependencies, and supports APIs such as Anthropic, OpenAI, and Gemini in proxy form.

The core idea is to block potential agent performance degradation during compression with statistical guardrails, achieving both cost savings and performance preservation at the same time.