Headroom cuts agent input tokens 60-95%, output barely moves

Headroom compresses what enters an AI agent's context, tool results, logs, RAG chunks, using AST- and schema-aware compressors, per the project's GitHub repo. A separate tool called Caveman compresses the other direction: the narration an agent writes back out, trimmed with prompt-level style constraints.

The split matters because tool-heavy agent loops burn most tokens on what tools return, not on what the model narrates back. Benchmarks posted with the two tools put input-side reductions at 60-95%, per Headroom's repo.

Output-side compression collapses to single digits on agentic coding runs by the same measure. That's the gap between an agent that runs out of context mid-task and one that doesn't. The repo doesn't break out results by compressor type or say which coding runs were tested.

The project's guidance follows that gap. Spend compression effort on the input side for tool-heavy agent loops. Save output-side trims for chat-shaped work, where the model's own text makes up most of what moves back and forth.