AI agents talk too much and read too much: caveman 4.0 splits into six rocks and cuts 63 tokens to 20
Synopsis
caveman 4.0 turns a single output-shortening tool into six separately switchable parts that are all on after install, aiming to make AI agents say less, read less, and run on less. It leaves prompts alone because Adobe found that compressing prompts makes answers worse, and it leaves code and tool calls alone because JetBrains found agent sessions are mostly code and tool calls. On ten developer questions, with "Be concise" as the real baseline, /caveman cuts 3% more at the median and /ultracave 35% more, and 63 tokens become 20 for the same answer.
Interpretation
caveman 4.0 becomes six parts instead of one output compressor, all enabled after install, each with its own switch; caveman status shows what is on and caveman off browse turns one off. Previously it only shrank what the agent says, and now it also shrinks what the agent reads and what it runs on. Two outside findings shaped the direction: Adobe found that cavemanning a prompt makes answers worse, so caveman never touches the prompt, and JetBrains found agent sessions are mostly code and tool calls, which the voice leaves alone.
The output compression keeps a fixed reply structure and shortens expression without breaking grammar, so answer quality is unchanged. "Be concise" is the real baseline, since new models already know it, and caveman compresses on top of it. Ten developer questions on claude-opus-5-5 show /caveman cutting 3% more at the median and /ultracave 35% more, with 63 tokens becoming 20 for the same answer, counted with tiktoken o200k.
The reading compression targets the files an agent reads, and the same answers stay correct. The whole session also carries the system prompt and tools, so it moves less than the files themselves. All 18 test files were answered correctly, with each file passed through the current compressor alone.
Perspective
The tool is aimed at developers running AI agents on their own machines, covering Claude, Codex, Gemini CLI, OpenCode and 30+ more agents, using the user's own keys and Claude Pro/Max login and requiring Node.js 22.13 or newer. It fits everyday coding sessions where shorter agent output, less read content, and lower running cost are wanted.
The output-compression percentages were measured on the 3.1.0 rules and have not yet been re-run on the new rules. The rules themselves cost about 1,160 input tokens, and if you pay per request, as with GitHub Copilot, shorter answers save nothing. The Claude app, phone, and Cowork paths have not been tested end to end.