AgentPlays
Sakana AI releases The Agentic Scientist (AI Scientist v3) as open code
Sakana AI released The Agentic Scientist, the third version of its AI Scientist, as Apache-2.0 code on GitHub (paper Oct 13, announced Oct 15). It wraps coding agents in a scaffold that proposes ideas, runs experiments and ablations, and drafts a paper from a task file and a scripted evaluator. Sakana frames it as co-discovery. In its Error Diffusion case study, human researchers set the question and rubric, chose directions, and checked results, references and claims before submission. Sakana says that paper was accepted at ALIFE 2026. The authors say researchers stay accountable for final claims. Agents run with full tool permissions, so Sakana recommends its Docker setup. (Sakana AI)
How you’d use it: If you want coding agents to propose ideas, run experiments against a scripted evaluator, and draft a paper from a task.md, try SakanaAI/AI-Scientist-v3 under Apache-2.0. Use their Docker setup: agents run with full tool permissions and need paid model API keys. Keep a human checking claims before anything goes out.