8 agents beat 1 by 2.3x in valid experiments per hour, six-hour local swarm benchmarks
cong_ml · x · 2026-10-27
Across three tasks (circle packing, job scheduling, bike demand prediction), eight agents averaged 2.3x more valid experiments per research hour than one, over six-hour runs twice per task. Final solutions were also better: 2.6279 vs 2.5689, makespan 2517 vs 2623, validation error 0.3754 vs 0.3905. Each agent gets rotating instructions — refine, try another algorithm, simplify, question an assumption — a simple version of AlphaEvolve-style prompt variety.
More from coding & agent
- figwig format ships with CLI tools and an agentic skill for bioinformatics sparse data — jmschreiber91 · 2026-10-28
- Harmonic launches Aristotle, an AI theorem prover that formally verifies C and Verilog code — xennygrimmato_ · 2026-10-28
- Designers argue personal agents must build user-fitting into the product, not leave it as a switch — sujingshen · 2026-10-28
- Nous Research Revamps Hermes Agent Plans: Cloud Agent Included, Daily/Weekly Limits Replace Monthly Credits — NousResearch · 2026-10-28
- Claude Code can now edit videos and cut TikTok-ready clips — 8 prompts to try — aitrendz_xyz · 2026-10-28
- Harmonic Launches Aristotle for Automated Formal Verification of C and Verilog Code — satnam6502 · 2026-10-28