A human-guided manager for AI research assistants
Original title: The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery
AI at workCoding with AIAI agents
Read by Charlie, a synthetic voice
The paper builds a manager program that coordinates code-writing helpers through research steps. Human researchers set direction and rules, check results, and edit the final output. The authors report an accepted paper co-created with this system and solo runs that found new methods, while noting the manager layer is key.
Our take: The evidence is mixed: an accepted paper with human help is real, but checks lean on an automatic reviewer, only one paper had outside review, and tests stay in computer research. It matters because the system still needs human rules and judgment.
- Left alone, AI helpers can miss broad ideas; human guidance matters.
- A manager program runs coding helpers in parallel and checks each stage against rules.
- The authors report an accepted paper with heavy human involvement and solo discoveries in computer research.
20 claims checked against the source Read the paper on arXiv PDF
Step by step
Start at the top. Each step adds one new idea. Stop wherever it makes sense.
The Agentic Scientist is a distributed pipeline for human researchers to conduct computational research with coding agents. It drafts tasks, generates ideas, plans sandboxed experiments, aggregates results, builds baselines and ablations, analyzes findings, writes reports, and applies rubric verification. A meta-controller manages agents and supports co-scientist or end-to-end autonomy.
-
I Helpers can miss the wider goal
Imagine a farm helper who digs only where the soil felt soft yesterday. The helper works hard. But it never asks if a better field lies over the hill. It just makes the nearest hole deeper. What has the helper missed? It misses choosing where to dig next. If people want new things, this keeps them close to old spots.
In real life: Farm helper = coding agent, a program that writes and tests code. Nearest hole = small local improvement. Better field = wider research idea. The paper calls this limited ideation and myopic hill-climbing, meaning it only improves nearby steps.
Takeaway: Workers using AI coding helpers may still need to set broad goals.
-
II Split research into checked stations
Now the farmer gives a row of stations. One station asks what crop could matter. One plants a small test plot. One compares plots. One writes what happened. The work is split so each step has a clear job. Why split the work? It makes hard thinking and testing visible.
In real life: Stations = distributed pipeline stages, spread across computers. Task drafting sets the task. Idea generation makes ideas. Planning chooses steps. Sandboxed experiments are isolated computer tests. Result aggregation combines outcomes. Baseline and ablation checks compare old ways and remove parts. Post-hoc analysis looks at results afterward. Report writing finishes the work.
Takeaway: AI research tools can be checked stage by stage before trusting claims.
-
III A manager enforces the checklist
A foreman now stands by the fence with a checklist. Before a basket moves on, the foreman tests it against the checklist. If it fails, it goes back. Several crews work different fields at once. The foreman sends help where the checklist says it is needed. Who decides when a step is good enough? The checklist does, with the foreman enforcing it.
In real life: Foreman = meta-controller, a manager program. Checklist = rubric verification, rule checks. Crews = parallel coding agents. Orchestration means coordinating many agents. Agents also self-verify by checking their own work. The same manager can direct different tools and environments.
Takeaway: Your AI tools may depend on manager rules as much as model quality.
-
IV Evidence is early and guided
The authors show two uses. End-to-end means from question to report. Co-scientist means guided by people. In one, people guide the system and submit a paper that was accepted. In another, it runs alone on machine-learning problems, where computers learn patterns. It also tests whether the manager matters. The authors say autonomy depends on both the base model and the coordination layer. This evidence is early. One accepted paper and the autonomous runs are not broad proof of full reliability.
Takeaway: AI-made claims may need human review before you trust them.
-
V Technical promise and clear limits
Technically, one discovery is a hypernetwork-based inductive bias, a built-in learning tendency, for rapid Transformer adaptation. A hypernetwork changes another network. Transformers are common pattern-finding models. Another result is a meta-trained, trained-to-learn evolution strategy, a learned trial-and-error search. It is compared with CMA-ES, a standard optimizer. The authors report matching or exceeding CMA-ES on five of seven unseen Brax control tasks. Brax simulates control tasks. Limits include fragile task drafting, limited literature-to-code baselines, and late literature search. Rubrics depend on evaluators. External peer review did not cover autonomous outputs. Evaluations stay in computational machine learning. Verdict: strong co-scientist orchestration; early for full autonomy.
Takeaway: This may affect research jobs; wider effects are background knowledge, unverified.
Checked against the source
Every statement in this brief had to match a word-for-word passage from the source; code checks each match, and a second AI model reviews the whole brief for errors. A brief that fails is not published. These are the 20 passages it rests on.
Show the 20 passages
Computer helpers can write code, launch experiments, analyze results, and operate for longer periods.
they write code, launch in silico experiments, analyze their results, and autonomously operate with ever-increasing time horizons.
Fully autonomous research agents can have a limited idea range and small-step search habits.
fully autonomous research agents suffer from a limited ideation radius and myopic hill-climbing search dynamics.
The system is an end-to-end pipeline for human researchers to conduct computational research alongside coding agents.
we introduce The Agentic Scientist, an end-to-end distributed pipeline for human researchers to conduct computational research alongside modern coding agents.
The system breaks research into steps such as task drafting, idea generation, planning, sandboxed experimentation, result aggregation, baseline and ablation construction, post-hoc analysis, and report writing.
our system decomposes research into task drafting, idea generation, planning, sandboxed experimentation, result aggregation, baseline and ablation construction, post-hoc analysis, and report writing.
Each step in the pipeline is powered by coding agents and leverages agentic self-verification.
Importantly, each step in the pipeline is powered by the next generation of coding agents and leverages agentic self-verification.
The overall research process is driven by a meta-controller that manages parallel research agents and enforces rubric verification at each stage.
The overall research process is driven by a meta-controller that manages parallel research agents and enforces rubric verification at each stage.
The same orchestration layer can be reused across different domains with different execution environments, coding harnesses, and evaluation metrics.
This design enables the same orchestration layer to be reused across heterogeneous domains with different execution environments, agentic coding harnesses, and evaluation metrics.
The system can be used as a co-scientist tool or to perform end-to-end scientific investigation.
Our system can be used across a spectrum of autonomy: either as a co-scientist tool, or to perform end-to-end scientific investigation.
In the co-scientist setting, a development version of the system was used to submit a single research paper.
Within the co-scientist setting, we leveraged a development version of our system to submit a single research paper
That paper was accepted for publication at ALIFE 2026.
which got accepted for publication at ALIFE 2026.
The paper was produced by human researchers using the system as a tool, not submitted as unaltered system output.
We emphasize that this single submitted ALIFE paper was produced by human researchers using The Agentic Scientist as a tool, and not submitted as an unaltered system output.
The system was applied fully autonomously to open problem-solving tasks in machine learning.
We further apply The Agentic Scientist fully autonomously to open problem-solving tasks in machine learning
The system demonstrated capability of scientific discovery across different domains of computational research.
and demonstrate the system's capability of scientific discovery across different domains of computational research.
One discovery was a hypernetwork-based inductive bias for rapid Transformer adaptation.
including: a hypernetwork-based inductive bias for rapid Transformer adaptation
One discovery was a meta-trained evolution strategy that matches or exceeds CMA-ES on 5 of 7 unseen Brax control tasks.
a meta-trained evolution strategy that matches or exceeds CMA-ES on 5 of 7 unseen Brax control tasks.
Ablations suggest that reliable scientific autonomy depends not only on base model capabilities but also on the cross-agent orchestration layer.
Our ablations suggest that reliable scientific autonomy depends not only on base model capabilities, but also on the cross-agent orchestration layer.
Verification is only as complete as its task-specific evaluators and rubrics.
the system's verification is only as complete as its task-specific evaluators and rubrics
The paper-level metric is an AI-reviewer proxy rather than full evidence of scientific correctness.
its paper-level metric is an AI-reviewer proxy rather than full evidence of scientific correctness
Only the Error Diffusion paper underwent external peer review.
only the Error Diffusion paper underwent external peer review.
Evaluations are restricted to computational machine learning.
Our evaluations are also restricted to computational machine learning