One of our autoresearch runs sat flat for 60 steps. One message got it moving again. Everyone is pushing research agents toward full autonomy, but what helps most is being able to step in when a run goes wrong, without breaking the loop. (1/6)

We made Weco autoresearch steerable: message a run mid-flight and it branch a new search, keeping everything it found. (2/6)

Hand it an idea from a paper. "Read this arxiv link and try it." It reads the paper and runs the idea as a new branch. (3/6)

Or start several directions at once, each with its own step budget. "Spend 10 steps trying rust, 10 trying mlx." They run in parallel from the best node. (4/6)

Every steer becomes its own subtree. You can see where each idea came from and which steer actually moved the metric, so you know what to build on and what to drop.

It works right away from a Claude Code or Codex session. We also let you connect your local Claude Code straight to the dashboard, so everything is together. Get started with "pipx install weco & weco setup claude-code"

More details:

turns out bots need babysitting too

Why can’t it find escape path by itself? Why still need human to point out arxiv?

Good q. Because human experts are usually still more creative, think deeper, or have global view that AI don’t have in its context window. For arxiv, it’s not about let it read papers in general but “this” paper

imo, if it stuck in local and requires a hand from human, that's smart hill-climbing but not autoresearch or long-horizon task. frontier intelligence should be able to fire well-composed search queries, find related info and escape from local optimum by itself.

Frontier intelligence still benefits a lot from human in the loop: It can escape from local optima but will still exhaust ideas eventually, because it still have a finite and small context window. On high level, human inputs are helpful because human intelligence is not a subset of AI

No that I agree. I understand the human inputs r valuable, but scarce like super very scarce especially when autoresearch on nontrivial question. In ur opinion, is the vision of autoresearch just be an intern and execute our thoughts of the know-unknown or should it focus on explore the unknown-unknown?

Yes, human input is scarce. At the same time tokens& experiments are also expensive, so the goal of the product is to improve research outcomes per unit of human effort. In the longer term the vision is the latter, explores unknown unknowns endlessly. I’m sure we will get there eventually, need consistent investment on research side to get there though

Poked around the docs. Looks very cool 👍

I think the world would be better if people like you were executed by the state