Sakana AI open-sources The Agentic Scientist, its AI Scientist v3 for coding agents

What Sakana AI released

Sakana AI has open-sourced The Agentic Scientist, the third generation of its AI Scientist line, as a Python pipeline in the SakanaAI/AI-Scientist-v3 repository under the Apache 2.0 license. The method is described in an 88-page arXiv paper, The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery, submitted on 13 October 2026 by Robert Tjarko Lange, Luca Grillotti, Yutaro Yamada, Rujikorn Charakorn and David Ha. Sakana's project page and the public repository were widely shared on 15 October, which is the date this article uses for the release.

The headline change from the first two AI Scientist systems is architectural. Instead of one model trying to write a whole paper in a single long conversation, the new system wraps off-the-shelf coding agents in a scaffold that controls what each agent sees, what it may write, how its work is scored and how results pass to the next stage. Sakana says its earlier systems, including work published in Nature, showed that much of the research loop could be automated but also produced underdeveloped ideas and buggy implementations. The Agentic Scientist is its attempt to make that loop more rigorous and evidence-grounded.

Alongside the pipeline, Sakana released AutoIllustrate, a companion tool that generates scientific diagrams from an editable draw.io source and then restyles them with an image model.

How the pipeline turns a task into a paper

The README lists six phases that run in order: task drafting, investigation, ablations, a sealed final evaluation, post-hoc analysis and paper writing, with optional review and revision phases afterwards. A run starts from a task directory holding a task.md description, an initial program and an evaluator script that scores attempts. Each phase runs in its own container and reads only a run-local snapshot of the task, so the shipped examples are never modified.

Investigation is where most of the search happens. According to Sakana's project page, a central controller splits a fixed budget of research cycles across parallel workers. Inside each worker an ideation agent proposes a hypothesis, an implementation plan and rubrics stating what the experiment must show; an execution agent then edits the candidate program, runs evaluations and records metrics. Workers share memory entries and distilled conclusions at cycle boundaries, and periodic context resets hand work to a fresh agent so the search can change direction without one ever-growing chat.

After search, the system picks the best correct program across workers and builds baselines and ablations around it. An analysis agent mines logged metrics for post-hoc evidence, and a writer produces a LaTeX paper with comparison plots and an explicit evidence map linking each claim to its source files. The README says the paper phase writes, compiles and audits the manuscript against that evidence, and the project page says automated checks and an independent audit look for unsupported numbers, inconsistent comparisons and overclaiming before a reviewer agent assesses presentation and framing.

The ALIFE 2026 paper and what the humans did

Sakana's main proof point is a single peer-reviewed paper. A development version of the system was used as a co-scientist on Error Diffusion, a biologically motivated learning rule, and the resulting paper, titled "Diffusing Blame: Task-Dependent Credit Assignment in Biologically Plausible Dual-Stream Networks," was accepted as an oral presentation at ALIFE 2026, according to the project page.

The project page is unusually specific about the human role. Researchers set the scientific question, the baseline and the evaluation rubric; moved the work from classification into reinforcement learning and prepared those environments; independently checked implementations, results and references; revised claims that exceeded the evidence; and folded reviewer feedback into the camera-ready version. In one early run, human inspection caught a high-scoring candidate that violated the intended learning rule, which prompted the team to tighten the rubrics. Sakana describes that as reward hacking, the failure mode where an agent satisfies the metric rather than the method.

Sakana says it submitted only this one paper for external review, and only after human verification. The other manuscripts the system produced were kept as internal evaluation artifacts. That framing matters: the result is one human-steered paper, not a batch of autonomous acceptances.

Two fully autonomous case studies

The paper and project page also show runs where the system worked fully autonomously from a goal and an existing codebase. In the first, Sakana pointed it at the public Doc-to-LoRA repository, which encodes document chunks as low-rank adapters for a frozen language model. The system found that merging adapters before seeing the question made facts interfere, and developed a way to select the relevant adapter based on the query. With refinements to text selection and chunk boundaries, Sakana reports exact-match accuracy rising from 52% to 98% without new trainable parameters or retraining. The project page notes those numbers come from synthetic multi-needle retrieval, and that generalization to natural documents remains untested.

In the second, the system explored learned evolution strategies anchored to CMA-ES, a standard black-box optimizer. The arXiv abstract says the resulting meta-trained strategy matches or exceeds CMA-ES on 5 of 7 unseen Brax control tasks, and the paper says the optimizer is now integrated in the evosax library. The system's own ablation was humbling: a control with the learned adjustments disabled kept much of the advantage and achieved better final returns in longer runs, with the learned part mainly speeding up early progress. Sakana presents that as the evidence-generating part of the loop working as intended.

Model versus harness: the 70-campaign comparison

To test how much the choice of model and agent harness matters, Sakana ran 70 research campaigns across five manually prepared tasks. The evaluation covered nine frontier models, four from OpenAI's GPT family, four Claude models and one DeepSeek model, across four harnesses: Codex, Claude Code, OpenCode and Pi, for 14 model-harness combinations. Each campaign got a common budget of 20 idea iterations, and the generated manuscripts were scored by an AI reviewer that Sakana treats as a proxy for paper quality rather than proof of correctness.

Newer models generally earned higher reviewer scores, especially within the Claude family, and the harness had a smaller effect than the underlying model. Pooling all 70 campaigns, Sakana reports a positive correlation of r = 0.44 between within-task numerical performance and reviewer scores. In a narrower comparison on an RL fine-tuning task, the paper says Pi reached scores similar to Claude faster and more consistently while cutting API costs by 75%. Those are the authors' own measurements on their own tasks; no outside replication has been published yet.

How to run it, and what it costs to try

The repository targets researchers with a program to improve and an evaluator that scores it. Setup needs Python 3.11 or later with uv, Docker, and API keys or a Claude login in a .env file. The recommended first run uses the claude agent inside Docker on a CPU-only circle-packing example: agsci doctor checks the host, agsci run examples/circpack --dry-run previews every command, and a trimmed run produces a compiled paper.pdf. A browser-based run inspector shows every phase.

The pipeline drives claude, codex, cursor, opencode, gemini and pi through Lange's Headless CLI, so the same experiment can be repeated across agents and models. A GPU preset for GRPO training needs two GPUs, and the README warns each evaluation trains for one to two hours. An optional review phase sends the finished manuscript to an OpenAI-backed reviewer. Drafting a task from a bare research brief is labeled experimental: it works for small tasks like circle packing but may fail on larger ones.

Running it means paying for model usage. Sakana does not publish a per-run price; the repository documents per-run cost output and includes a separate agent-billing guide for the supported agents.

Security warnings and stated limitations

The README carries a direct security note. The Agentic Scientist launches coding agents with full tool permissions, so they can run arbitrary commands, inherit the API keys they are given and be steered by the task files and papers they read. Sakana asks users to keep runs inside the Docker image, run only trusted tasks and sources, and review run directories before sharing them because agent logs can echo document contents and paths.

The project page lists research limits too. The system searches literature only while writing the manuscript, not during ideation, so new hypotheses do not build directly on prior work. It lacks a reliable way to turn published methods into state-of-the-art baselines, so most comparisons are variants of its own winning program. All tasks so far are computational machine learning, leaving fields with physical experiments or human subjects untested. On mature benchmarks, Sakana says search can collapse into hyperparameter tuning. AutoIllustrate's benchmark gains rely on a single vision-language model as judge.

Where it fits in the autoresearch wave

The release lands in a busy week for agent-run research. ByteDance and Duke described how four Claude agents post-trained Qwen3-4B to instruct level with Athanor, and Google Research open-sourced Harness² for recursive agent improvement. Microsoft researchers have argued that clearing an agent's chat history helps long research runs escape ruts, an idea that echoes The Agentic Scientist's periodic context resets.

What sets Sakana's release apart is the full research envelope: sealed evaluation, ablations and an evidence-linked paper in one open, reusable package, plus a peer-reviewed paper whose human edits are documented in the arXiv appendix. The open questions are whether outside groups can reproduce the Doc-to-LoRA and CMA-ES results, how often rubric checks miss shortcuts without human review, and whether conferences will ask for disclosure of agent contributions as Sakana itself recommends.