4 Comments
User's avatar
Andrei Danilchenko's avatar

Nice test for agents!

Curious have you tried any version of autoresearch for this problem? Or swarm approach was actually the autoresearch?

With proper validation autoresearch can bring more strict harness (you just make one step at a time and maintain log of all experiments) and much better overall result

Aleksey Tikhonov's avatar

In this case I used /workflows -- a pretty recent CC feature (https://claude.com/blog/introducing-dynamic-workflows-in-claude-code) for swarming control.

I also tried ShinkaEvolve (Auto-Evolve analogue from Sakana), but it was less successful. Autoresearch seems a bit too simple for this.

Andrei Danilchenko's avatar

What was generated workflow? Did it have some loop for idea iterations?

Aleksey Tikhonov's avatar

Nope. It's straightforward, 6 separate parallel agents to research specific aspects; 3 parallel agents to validate candidates from the first 6; 1 last agent to merge proper ideas into final solution.