Weco AI

What kinds of problems can it tackle?

Anything you can score with a script, and whose evaluation cost is manageable, should be considered. We've seen success on tabular ML feature engineering, prompt optimization, video encoding parameter tuning, frontend bundle size, backend latency, and RL reward design.

What kind of improvement should I expect?

Most users run it for a few dozen or a few hundred steps. Best-score improvement scales roughly as n^(1/7) for a bounded metric in [0, 1], showing diminishing returns, but you can still push further if the metric truly matters. Sometimes users push the score above that scaling law with updated or additional instructions (e.g., providing a list of novel ideas).

Can Weco optimize across multiple files or multi-prompt systems?

Yes. Point it at a directory or a set of files and Weco coordinates edits across them.

Can I use my own LLM API keys (BYOK)?

Yes. With BYOK, optimization LLM calls go through your OpenAI, Anthropic, or Gemini keys, billed directly by your provider. A small amount of platform credit is still used for Weco's own orchestration.

Where does my code run? Is my data sent anywhere?

Your code and data stay on your machine, and every evaluation runs against the source you control. The Weco agent runs in our cloud: it sees the diffs it proposes and the metric values your eval prints, but not your training data, model weights, or private inputs. Your code isn't used to train models.

Do I own the generated code?

Yes. You keep full intellectual property rights. Weco writes plain text into your repo, under your license, with no hidden terms or dependencies.

How much does it cost?

We offer 20 free credits (~100 optimization steps) to start. Each step costs a fraction of a cent plus whatever LLM tokens the proposed diff uses. Typical small-scale runs land around $10–$30. You can also bring your own API keys to route LLM spend through your provider directly.

How long does it take?

It can range from a few minutes to days, depending on how long the eval takes. You can decide the time budget by setting the number of steps. Roughly (eval time + LLM generation time) × steps, and the LLM generation time is usually around 1 to 3 minutes.