La Agente Óptima: Towards Agentic Self-Driving Laboratories
Marcel Müller, Jiaru Bai, Willi Gottstein, Abhijoy Mandal, Mohammad Nazeri, Elia Savino, Yanlin Fang, Sujoy Das, et al.
cs.AI · cs.MA · physics.chem-ph
Quick verdict
Agentic framework decouples LLM reasoning from persistent BO campaigns, shown on benchmarks and two live labs; strong engineering but thin statistics.
Summary
The paper introduces La Agente Óptima, an agentic framework in which an LLM coordinating agent and a specialist subagent author, execute, and supervise Bayesian optimization (BO) campaigns whose state persists in a separate service (BO-MCP, backed by BayBE or BoTorch). By keeping repetitive loops in programmatic campaign code and returning control to the agent only for interpretation or revision, the system targets accessible, auditable self-driving-laboratory operation. Evaluation spans ablations against three architectures on Ackley 6D and the Shields arylation benchmark (three 60-attempt runs each), eight specialist LLMs, four computational discovery campaigns, and two physical platforms. The proposed architecture achieved the best Ackley performance (final 0.770; MBSF@60 0.442) and reached maximum Shields yield in all three runs, though Main-script attained a higher Shields MBSF@60 (0.891) at roughly half the cost. On RAISE, Óptima diagnosed a flawed failure-encoding scheme that biased the surrogate, and moved contact angle from 71.4° to 67.8° without reaching the 65°±1° target. Over a five-day RoboChem-Flex campaign, yield rose from 30.0% to 58.8% in 23 experiments; compared with a human-directed campaign (70.9% in 50 experiments; 63.9% at matched consumables), the agent run cost 13% less and used less than half the starting material. Limitations include small run counts (n=3) without significance testing, mixed ablation evidence, proprietary frontier-model dependence, and physically narrow single-run demonstrations.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Separating LLM reasoning from programmatically executed campaigns with persistent optimizer state (BO-MCP) makes Bayesian optimization accessible, consistent, and auditable for domain scientists.
- The Specialist-script architecture achieves the best Ackley 6D performance (final 0.770, MBSF@60 0.442) and reaches maximum Shields arylation yield in all three runs, though Main-script attains higher Shields MBSF@60 (0.891) at about half the cost.
- Óptima sustained closed-loop optimization on two physical platforms, raising flow-chemistry yield from 30.0% to 58.8% over 23 experiments and contact angle from 71.4° to 67.8°, while autonomously diagnosing flawed failure-encoding and a zero-yield analytical anomaly.
- Relative to a human-directed campaign on the same transformation, the agent campaign cost 13% less, used less than half the starting material, and selected a more mass-efficient operating point (58.8% vs 70.9% full / 63.9% at matched consumables).
Strengths
Honest boundary and negative results
The RAISE campaign missed its 65°±1° target and is reported as such, with the agent inferring infeasibility; the Pollice best-so-far trace is described as almost flat with the 0.227→0.223 eV gain attributed to a lucky seed; TD-DFT evaluator limits are quantified (Spearman ρ=0.79, Pearson r=0.89, +0.33 eV offset, no inverted gaps reproduced); the Co campaign's 6-feasible-evaluation effective budget is called too small by the authors themselves (SI C.3.3).
Rigorous equal-budget ablation protocol
Appendix B uses a global 60-attempt budget that counts smoke-test and debug evaluations, prespecified validity criteria (SI B.1.1), and reports protocol failures as outcomes: Local-BO met the 90%-yield Shields criterion in only 1/3 runs, and 8/48 specialist-model failures are itemized (seven global-budget, one ownership-only).
Real closed-loop operation with failure forensics
State-of-the-art provenance and cost accounting
Agent-driven search-space self-correction
Weaknesses
No inferential statistics
Mixed ablation evidence for the headline design
Confounded human comparison
Proprietary-model dependence and unclear code release
Digital campaigns are re-demonstrations with modest gains
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a systems/method paper from the Aspuru-Guzik (Toronto) and Noël (Amsterdam) groups presenting **La Agente Óptima**, an agentic framework in which an LLM orchestrator and a BO-specialist subagent author, execute, and supervise Bayesian-optimization (BO) campaigns. Optimizer state persists in a dedicated **BO-MCP** service (MCP + REST; BayBE default backend, BoTorch legacy backend with TuRBO auto-dispatch at 20+ parameters), and repetitive loops run as generated Python campaign code rather than per-iteration LLM tool calls (Sec. 2.1–2.2). Central claims: (1) the reasoning/execution/state separation makes long-running BO accessible, consistent, and auditable; (2) the design outperforms architectural ablations; (3) it operates end-to-end on two live platforms — RAISE (contact angle) and RoboChem-Flex (flow photochemistry) — after four computational discovery showcases.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is La Agente Óptima?
An agentic framework in which LLM agents (an orchestrator plus a specialist subagent) construct and supervise Bayesian-optimization campaigns for self-driving laboratories. Optimizer state lives in a persistent service (BO-MCP, backed by BayBE or BoTorch), so repetitive loops run as generated Python code and the LLM intervenes only for interpretation or campaign revision, keeping decisions auditable.
What did it achieve on real lab hardware?
On the RAISE platform it moved a contact angle from 71.4° to 67.8° without reaching the 65°±1° target, diagnosing a flawed failure-encoding (failed measurements recorded as 180°) along the way. On RoboChem-Flex it ran a five-day flow-photochemistry campaign raising yield from 30.0% to 58.8% over 23 valid experiments, at 13% lower cost and less than half the starting material of a human-directed campaign that reached 70.9%.
How is it different from just asking an LLM to optimize?
The paper's ablations show that a pure LLM tool-loop and locally coded BO both underperform the proposed design on Ackley 6D; programmatic campaign execution with an established BO backend and persistent state was more reliable — the Local-BO ablation reached the 90%-yield criterion on the Shields benchmark in only 1 of 3 runs.
Is the code or data available?
The provided text describes extensive reproducibility artifacts (seeded campaigns, BO-MCP exports, SHA-256 bundles) and detailed SI documentation, but shows no public code-release statement; headline results depend on proprietary GPT-5.x-class models and internal services, so full external replication is not currently possible from the paper alone. The Shields arylation benchmark itself derives from a published public dataset.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with [...sampling gap...] markers; middle sections (parts of Sec. 3.1.1, the end of Sec. 3.3.4, SI B.1.1/B.1.2 details, SI C.5/C.6, and any safety or discussion sections) were skipped and may contain relevant content, including a code/data availability statement.
- codeAvailable=false reflects that no release statement appears in the provided excerpt, not a verified absence from the full paper.
- The full validity criteria of SI B.1.1 are cited from their summary in the main text; the SI section itself was not in the provided excerpt.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.