This repository collects the raw data underlying La Agente Óptima: Towards Agentic Self-Driving Laboratories — the run evidence, audit metadata, plotting code, and figures for the framework comparison, plus the raw run evidence for the showcase campaigns reported in the main text and SI.
La Agente Óptima (Óptima) is the agentic framework evaluated in the paper: a coordinating agent plus a Bayesian optimization (BO) specialist subagent that authors and supervises BO campaigns, keeping LLM reasoning separate from the programmatically executed optimization loop.
BO-MCP is the BO backend service Óptima optimizes through. It holds campaign configuration, accumulated observations, and the per-campaign action ledger in its database, and exposes them over MCP and a REST API. Repo: AccelerationConsortium/bo-mcp.
snapshots/— the framework comparison (SI, Framework comparison): raw evaluation cells, controller metadata, and the comparison report. Described below.showcases/— the showcase campaigns reported in the paper and SI: OSL combinatorial discovery, four finite-space computational campaigns, the RAISE closed loop, and the RoboChem-Flex flow-photochemistry campaign. Seeshowcases/README.md.
Two showcases ran Óptima against external self-driving laboratories:
- RAISE (Robotic Autonomous Imaging Surface Evaluator) — closed-loop contact-angle formulation discovery. Repo: Frank-Gu-Lab/RAISE. Paper: Nazeri et al., Digital Discovery 5, 2254–2270 (2026), 10.1039/D5DD00531K.
- RoboChem-Flex — modular, low-cost flow-chemistry platform used for the multi-objective flow-photochemistry campaign. Repo: Noel-Research-Group/Robochem_Flex. Paper: Pilon et al., Nature Synthesis (2026), 10.1038/s44160-026-01053-0.
The comparison evaluates the proposed Óptima architecture
(Specialist-script) against the Main-script, Tool-loop, and Local-BO
ablations, and additionally varies the LLM assigned to the BO specialist, over
the Ackley 6D and Shields arylation benchmarks.
The current snapshot includes the corrected BayBE duplicate-resuggestion replacements and the latest comparison report. Results remain subject to manuscript review; prior snapshot states are preserved in Git history.
snapshots/2026-08-08-current/ contains:
raw/: all 66 canonical cells used by the current report, organized by arm, case, and repeat. Cell directories retain conversations, metrics, logs, scripts, result artifacts, and request ledgers where available.source_controls/: controller metadata and PostgreSQL snapshots used to reconstruct campaign-level evidence.report/: the full comparison report, figures, pricing rules, cost and budget audits, selection audit, plotting scripts, and integrity manifest.
The published copies have been scanned and sanitized for API-key-shaped values.
Original research-server evidence was not modified. Sanitization counts are in
snapshots/2026-08-07-preliminary/SANITIZATION_REPORT.json.
MANIFEST.sha256 is the portable integrity manifest for this repository. The
manifest under report/control/ is retained from the research server for source
provenance and contains its original absolute paths.
Create an environment with Python 3.12 and install the plotting requirements:
python -m pip install -r requirements.txt
python scripts/rebuild_figures.pyThe script reads the frozen REPORT_DATA.json and writes the figures back to
the snapshot's report/figures/ directory.
- Budget PASS: exactly 60 attempted objective evaluations in total across all campaigns owned by the run.
- Scientific PASS: the evaluation results are complete, unique, schema-valid, and use the expected backend.
- Architecture PASS: the intended main-agent/specialist ownership and required artifacts are present.
Campaign count is descriptive. One or multiple campaigns are allowed.