Sakana AI Releases The Agentic Scientist, Paper Accepted at ALIFE

Sakana AI (David Ha's team) has released the third generation of the "AI Scientist"—The Agentic Scientist. The key change is a shift from "a single agent writing papers" to "parallel agents orchestrating the full research loop." The system lets human researchers collaborate with modern coding agents on computational science research, and a paper it completed autonomously has passed human peer review at ALIFE 2026.

Confirmed

• The system breaks research into four stages: task setup, parallel idea investigation (with rubric-based self-verification), post-hoc ablations and evidence generation through analysis, and finally producing a complete paper; multiple CLI agents collaborate in a distributed fashion.
• Alongside the paper, the team open-sourced the codebase plus AutoIllustrate: a drawio-based agentic pipeline that turns a paper's methods abstract into editable vector .drawio diagrams.
• The paper includes multiple ablations and real-world case studies showing different autonomy levels, including an extension of tan51616's Doc-to-LoRA work where the system autonomously discovered a query-based multi-LoRA merging strategy, as well as extended experiments on evolutionary strategies.
• RobertTLange traced the team's evolution in "automating AI research with LLMs": they explored this with GPT-4 back in 2024 (discovery work), and this release is a systematized upgrade.

Why it matters

• The system demonstrates the feasibility of coding agents across the full research pipeline—from topic selection to passing human peer review—and is seen as a landmark advance for "AI automating AI research."
• Editable drawio diagram output and the open-source release lower the barrier for other researchers to reuse and audit the work, letting the community build their own automated research workflows on top of the pipeline.

2026-10-15 ~ 2026-10-15 · 6 related posts

• [source] Sakana AI's Agentic Scientist orchestrates coding agents for research, lands ALIFE 2026 paper — omarsar0 · 2026-10-15
• SakanaAI's Agentic Scientist Writes Paper That Passed Human Peer Review at ALIFE — RobertTLange · 2026-10-15
• Agentic Scientist Case Studies: Auto-Discovered LoRA Merging and ES Extensions — RobertTLange · 2026-10-15
• SakanaAI Open-Sources AutoIllustrate: Editable draw.io Paper Figures from Method Text — RobertTLange · 2026-10-15
• [source] Sakana AI releases Agentic Scientist plus AutoIllustrate pipeline for editable paper figures — RobertTLange · 2026-10-15

1 near-duplicate retellings: hardmaru