How the swarm collaborates: shared best program plus rotating instruction variety

cong_ml · x · 2026-10-27

The eight agents collaborate through one shared artifact: the best program so far. When an agent beats it, others start from the new best. But agents starting from the same program tend to try the same things, so each rotates through distinct instructions — refine a part, try another algorithm, simplify, question an assumption — a lightweight take on AlphaEvolve-style prompt variety.

Related event: cong_ml runs 8-agent research swarm on a single 48GB M4 Pro with one local 40B model(7 posts)→

Original post →

More from coding & agent

coding & agent channel →