Weco

Weco is built by Weco AI. Its weco command-line tool runs LLM-guided tree search over your code, executes your evaluation script, and proposes new versions based on the metric you choose. It is distributed as an Apache-2.0 open-source CLI and works with a hosted dashboard.

What It Is

Weco is an autoresearch tool for optimizing anything with a measurable objective. You point it at one or more source files, give it an evaluation command that prints a metric (such as accuracy, speedup, loss, or cost), and set a goal to maximize or minimize. Weco then iterates: it edits the code, runs the eval, parses the result, and uses the outcome to guide the next change. When the run finishes, it can apply the best version to your source file.

Documented applications include agent and harness engineering, prompt engineering, ML model development, and performance work such as CUDA and Triton GPU kernels.

How a Run Works

• weco run takes --source or --sources for multi-file optimization, an --eval-command, a --metric, a --goal, and a step count.
• Models from OpenAI, Anthropic, Google, and hosted open models can be selected, and you can bring your own provider keys.
• Runs can be inspected and managed with subcommands such as status, results, diff, stop, and instruct for mid-run guidance, and interrupted runs can be resumed.
• A review mode lets you inspect, edit, and submit nodes yourself.
• Optional logs save per-step output and code snapshots.

Coding Agent Skills and Observe

Weco installs as a skill for Claude Code, Cursor, Codex, and OpenClaw through weco setup, so an assistant can write the evaluation, configure the run, and explain the results. weco observe is a separate mode for tracking experiments from your own agent or script in the Weco dashboard, with tree visualization, code diffs, and metric tracking, without handing over control of the optimization. A Python SDK is also available.

Company Context

The project's site describes Weco's research focus on recursively self-improving AI and cites its AIDE algorithm. The platform page states the dashboard shows the search as an observable tree and that runs can be steered mid-run.