Recursive Autonomous Science

RAS lets an agent improve its own harness from new experiments, giving small models frontier performance on challenging scientific tasks at 1/30 the cost.

Written by Yibo Wen on November 3, 2026

PaperCodeAutonomous Science Bench BibTeX

Runs resolved by domain

Base
Flat
RAS

GPT-5.6 Luna · Codex

Sonnet 5.5 · Claude Code

Resolution rate vs. cost per run

Leaderboard and RAS · per run

Leaderboard frontier
RAS frontier

A scientific agent is a frozen model wrapped in a harness, the prompts, memory, skills, and scripts that shape how it plans experiments and turns results into a predictive model. RAS keeps the model fixed and lets the agent rewrite its own harness, guided by how its campaigns on Autonomous Science Bench are scored. A campaign passes only if the experiments the agent chose clear a discovery gate and the model it delivered clears a learning gate. The animation below follows one step of that loop.

Step

k = 1, \dots, K

one model runs the campaign and edits its harnessk = 2

Verifier feedbackruns of

H^{\star}

H_1

Discovery91%bottleneckLearning106%bottleneck

Propose

Planning editorthe experiments

Modeling editorbottleneckthe delivered model

Screen

H_k^{\mathrm{plan}}

× 2 runs

H_k^{\mathrm{mod}}

replayed, no runspromising

Confirm

the most promising branch

H_k

× 2 fresh runs

or keeps

H^{\star}

Select

gain > 1 SEin mean score

S

· +0.22

no gate lostthat

H^{\star}

passes

kept:

H^{\star} \leftarrow H_1

the next step edits the best harness so far,

H^{\star}

discovery passed learning passed still to runIllustrative runs and scores

Start the next step. The kept branch becomes $H^{\star}$ . Learning now clears its threshold, so the next bottleneck is discovery.

Overview

Many of the ways small models fail these campaigns are procedural, in how they spend their budget of experiments and how they turn measurements into a model. Those procedures live in the harness, so a better harness can close the gap to frontier models without a stronger model. RAS finds that harness in three ways.

Guided by the science
The weaker gate, discovery or learning, decides what each step edits.
Efficient with evidence
Replay and simulation screen edits without new runs; fresh runs confirm a gain.
Frontier at 1/30 the cost
With RAS, Luna and Sonnet match or beat Opus 5.5 and GPT-6 Astra.

Method

Each recursion begins by running the base harness H₀ three times. A step then reads the verdict of every run and finds the bottleneck, the gate that falls furthest short of its threshold. The editor for that stage proposes changes to the harness, and RAS screens each one before spending new runs. A modeling change is replayed on runs that already exist; a planning change is simulated or run twice.

The most promising change is confirmed on two fresh runs. It is kept only if its gain in mean score exceeds one standard error, the run-to-run noise, and it loses no gate the current harness passes; otherwise the step rolls back. After the last step, the selected harness is evaluated on six fresh runs.

Discovery gate

Planning editor

Edits: The instructions, and the scripts, skills, and subagents that plan the campaign.
Screened by: A model-guided planner simulated through the campaign’s lab, or two runs.

Learning gate

Modeling editor

Edits: modeling.md and the modeling procedure in scripts/model/.
Screened by: Replay on earlier runs, scored by the campaign’s own verifier, at no run cost.

The editors work only from verdicts, so they can improve how the agent works but cannot hand it the answers. Five safeguards hold that line.

Kept honest

Verdicts, not answers: editors never see the verifier’s labels, unmeasured outcomes, or the lab’s reference files.
A closed sandbox: editors reach only the model APIs.
An answer guard: rejects a harness that names a design, sequence, composition, URL, or prebuilt predictor.
A procedure audit: rejects a modeling procedure that still ranks the evidence after its targets are shuffled.
Held-out evaluation: evaluation runs never help choose the harness.

The Harness

A harness is a small tree of files that RAS mounts on a frozen executor, Codex or Claude Code. Some files are read on every call, some only when the agent invokes them, and some run as scripts. One script, the modeling procedure, fits the delivered model after the run, which is why a modeling change can be scored by replaying old runs.

Harness

H

the files RAS edits

Contextevery call

System promptsystem_prompt.md
Guardrailsguardrails.md
Memorymemory.md
Tool guidancetool_guidance.md
Delivered modelmodeling.md
Campaign notescampaign_notes.md

On demandwhen invoked

Skillsskills/<name>/SKILL.md
Subagentsagents/<name>.md

Executedfrom the shell

Scriptsscripts/<name>.py
Modeling procedurescripts/model/fit.py

Frozen executor

Codex or Claude Code, with the context levers in AGENTS.md or CLAUDE.md, and the directories mounted at /app/harness.

With $H_0$ every file holds a placeholder and nothing is mounted, so Base runs the stock agent.

Lab serviceasb

One batch per round, measurements back, every job in the ledger.

Ledger

final_model.js

The output of fit.py on the run’s own rows.

Verifierofficial

discovery on the ledgerlearning on the model

Score both gates. The verifier scores discovery on the ledger and learning on the model. A modeling edit changes only fit.py, so replay can score it.

RAS edits only the harness files; the executor, lab, and verifier stay fixed.

Ten files, one harness

In H₀ every file is a placeholder, so the agent behaves like the stock model. The editors fill these files in over the recursion. Pick one to see it before and after an illustrative edit.

Context · every call

On demand · when invoked

Executed · from the shell

C6every callPlanning

Context

Campaign notes

Notes on this campaign’s rounds, batches, and budget.

campaign_notes.md

# Campaign notes
- Hold back part of the budget until a first peak is measured.
- Search for a second peak before committing the last rounds.# Campaign notes
- Hold back part of the budget until a first peak is measured.
- Search for a second peak before committing the last rounds.

Illustrative edit · Planning editorready

Results

On the six campaigns, RAS raises the share of evaluation runs resolved from 19% to 67% for GPT-5.6 Luna and from 19% to 86% for Sonnet 5.5. A non-recursive control with the same number of steps reaches 33% and 64%. Both RAS harnesses match or beat Opus 5.5 (67%) and GPT-6 Astra (72%) while costing $0.26 and $0.32 per run, against $9.61 and $10.57.

Over the recursion

The score S measures how close a campaign comes to passing, taking the weaker of its two gates, so 1 is a pass. Over the recursion, RAS raises the mean score by 0.20 for Luna and 0.36 for Sonnet, against 0.13 and 0.31 for the control, and only RAS lifts Luna’s mean past a pass.

Mean score by step

Flat
RAS
Pass

GPT-5.6 Luna · Codex

Sonnet 5.5 · Claude Code

Mean over six campaigns of S, the lower of discovery and learning as a share of threshold, so 1 is a pass. Step 4 is the evaluation runs.

Per campaign

With RAS, Sonnet 5.5 resolves every evaluation run in four of the six campaigns. Propylene remains the hardest for both models, at one and two runs of six.

Campaign	GPT-5.6 LunaCodex			Sonnet 5.5Claude Code
Campaign	Base	Flat	RAS	Base	Flat	RAS
Nuclease
Biosensor
Propylene
Suzuki
OER
Coercivity
Total	7/36	12/36	24/36	7/36	23/36	31/36

Evaluation runs that pass both gates, out of 6 per campaign.

Citation

If you use RAS, please cite it and Autonomous Science Bench.

BibTeX

@misc{wen2026ras,
  title={Recursive Autonomous Science through Self-Improving Agent Harnessing},
  author={The Autonomous Science Bench Team},
  year={2026}
}

@article{wen2026autonomous,
  title={Autonomous Science Bench: Evaluating AI Agents for Autonomous Scientific Discovery},
  author={Yibo Wen and Han Liu},
  journal={arXiv preprint arXiv:2610.18449},
  year={2026},
  url={https://arxiv.org/abs/2610.18449}
}

Runs resolved by domain

GPT-5.6 Luna · Codex

Sonnet 5.5 · Claude Code

Resolution rate vs. cost per run

Leaderboard and RAS · per run

Overview#

Method#

Discovery gate

Learning gate

Kept honest

The Harness#

Ten files, one harness

Campaign notes

Results#

Over the recursion

Mean score by step

GPT-5.6 Luna · Codex

Sonnet 5.5 · Claude Code

Per campaign

Citation#

Overview

Method

The Harness

Results

Citation