Instead of plugins, the agent writes down improvement ideas, edits its own Rust source and recompiles itself. It runs on DeepSeek 4 Flash on two local GPUs at about 138 tokens a second.
Result About 138 tokens a second on local GPUs
“The source code is in a fixed place so it can just rewrite and build the agent itself.”
Spotted: found in the wild. Details are as reported by the source above and have not been independently verified.
Had Claude Code build a coding agent in Rust, borrowing ideas such as Dirac’s line hashing for edits. Instead of plugins, the agent writes improvement ideas to a separate file, then updates its own source code, which sits in a fixed place, and recompiles itself.
It runs on DeepSeek 4 Flash on two RTX 6000 Pro cards at about 138 tokens a second. The developer says they openly borrowed from Pi, Dirac and OpenCode.
Build me a minimal terminal coding agent in Rust. It needs a chat loop with [model API] and tools to read files, edit files (using line hashes to target edits), run shell commands and search. Keep the source in [path]. Add a command that lets the agent write improvement ideas to IMPROVEMENTS.md, edit its own source and rebuild itself.
A starting prompt written by Agents in the Wild. Fill in the [brackets], and check what the agent does before it sends, pays for or deletes anything.