We got tired of tweaking prompts by hand and calling it optimization

The honest version of how most prompt work happens: you change a sentence, run it a couple times, it feels better, you move on. Maybe you're right. Maybe you just got a lucky sample and you'll never know, because nobody logged the comparison.

Autoresearcher is our attempt to stop doing that. You give an agent a goal, a skill to improve, a rubric to score against, and a budget for how many experiments it's allowed to run. It proposes one change at a time, runs the challenge, compares the new score against the current best version, and writes down what it decided and why.

You still set the direction — the agent just does the repetitions

This isn't "let the AI figure out what good means." You write the research program: the goal, the strategies worth trying, and the constraints — keep the prompt under a certain length, change one thing per experiment, whatever matters for your case. The agent's job is to grind through that search faster than you would by hand, not to invent the goal.

Every experiment leaves a paper trail

The log records the hypothesis, what actually changed, the score, the delta from the previous best, and whether it got kept or thrown out. Watch the score chart over a run and you'll see either steady improvement or a flat line telling you the current approach is tapped out — both are useful to know.

The rubric is doing the real work here

Autoresearcher doesn't decide what a good answer looks like. Your rubric does. If your rubric is vague, the results will be too — this tool makes that gap obvious fast, which is its own kind of useful.

Try it on a challenge you already know

Autoresearcher is live on selected Versalist challenges right now. Pick a skill you've already been hand-tuning, set a small budget, and read the experiment log when it's done. You'll probably learn something about your own prompt that you didn't know.