Compressor V2: three compression layers for a 50% LLM agent cost cut
7 hours ago
- #Token Compression
- #Coding Agents
- #Edgee Compressor V2
- Token compression is critical for coding agents due to economic pressure and operational bottlenecks like dollar cost, latency, context window limits, and throughput.
- Edgee's Compressor V2 includes three orthogonal strategies: brevity (reducing output tokens), tool surface reduction (TSR) for MCP tool catalogs, and tool result trimming for verbose tool outputs.
- Statistical analysis using paired sign tests, bootstrap confidence intervals, and coefficient of variation shows significant savings: brevity achieves ~30% cost reduction on SWE-bench Lite, TSR reduces tokens by ~33%, and tool result trimming adds incremental 5–10% savings.
- Brevity works by eliminating meta-text in outputs, preserving cache amortization, and targeting the most expensive token class without prefix changes.
- TSR rewrites tool catalogs into a single virtual tool, reducing prefix bloat and token volume, especially beneficial for tool-heavy workloads.
- The three strategies are composable and target different layers (output, prefix, history), allowing users to enable combinations per API key based on workload shape.