Sakana AI open-sources AI-Scientist-v3, orchestrating coding agents to automate scientific discovery
cong_ml · x · 2026-10-16
Sakana AI released AI-Scientist-v3 ("The Agentic Scientist"), an open-source (Apache-2.0) system that orchestrates coding agents for automated scientific (co-)discovery, with a blog post and arXiv paper.
Starting from a task description, agents propose ideas, run experiments, evaluate against a rubric, and iterate toward stronger programs, then run baselines and ablations, a sealed final evaluation, post-hoc analyses, and write a paper. The pipeline has four stages: Task Setup (task agent + evaluation function produce an initial program), Parallel Investigation (ideation and execution workers loop analysis→rubrics→refine over shared memory, with CLI agents in the pool), Evidence Generation (parallel execution agents, multi-seed metric comparisons, winner analysis), and Paper Writing (planning→writing→review with repair loop).
The authors frame it as a step beyond hill-climbing: illuminating unknown unknowns requires systems that build on prior discoveries, scale open-endedly, and support human co-discovery with evidence-grounded communication.
More from coding & agent
- Terminal Night 3 talk slides released, missing demo video to follow — 4310sy · 2026-10-16
- Agent developers are sitting on unused API credits that can exceed actual usage — EnvironmentalDay8864 · 2026-10-16
- Notion's bet on agents: abstract away deployment so users and agents just write code — simonguozirui · 2026-10-16
- Four Gemini 4 Argon agents beat Minecraft's Ender Dragon autonomously in 27 hours — Maxence Faldor · 2026-10-16
- Tencent's Terminal-RSI builds recursively self-improving terminal agents, hitting 19.84% on Terminal-Bench 4.0 — tencent · 2026-10-16
- Tsinghua's AutoSFT uses multi-agent collaboration to automate SFT data research — TsinghuaNLP · 2026-10-16