Weco: Production-Grade Autoresearch
A command-line tool that uses a tree-search approach guided by an LLM to iteratively rewrite code toward a measurable goal, such as optimizing an agent harness, a prompt, a training pipeline or a GPU kernel.
Weco is a CLI that optimizes code against a metric a user defines, rather than editing code by hand. It applies a tree-search approach guided by large language models: it proposes a change, runs the user's evaluation script, parses the result, and proposes the next change based on what worked, iterating toward goals such as accuracy, win rate, cost, latency, throughput or memory bandwidth.
Example uses include tuning the harness around an LLM agent's prompts and tools, refining or compressing prompts while holding accuracy steady, tuning feature transformations, architectures or a whole training pipeline, and speeding up code such as GPU kernels written in CUDA or Triton. It also installs as a skill for Claude Code and Cursor, so a user can describe an optimization goal in plain language and let the assistant configure and run Weco, monitor iterations and explain the results.
A separate observe command tracks an already-running optimization loop, whether driven by an LLM agent, a custom script or a manual workflow, in the Weco dashboard with a tree visualization, code diffs and metric tracking, without taking over control of the loop itself.