| name | scientific-discovery-agents-2026 |
| description | Design, evaluate, and operate agentic systems for biomedical and scientific discovery. Use when building or selecting agents for hypothesis generation, experiment planning, autonomous notebook analysis, lab-in-the-loop validation, pathology concept discovery, or multi-agent research workflows. |
| keywords | ["scientific-discovery","biomedical-agents","hypothesis-generation","lab-in-the-loop","multi-agent","evaluation"] |
| measurable_outcome | Produce an evidence-grounded scientific agent plan with autonomy class, data/tool boundary, human checkpoint, validation endpoint, and rollback criteria within 2 hours. |
| metadata | {"author":"Biomedical OS Team","version":"2026.05"} |
| source_reliability | [{"source":"primary_literature","score":1,"rationale":"Workflow is grounded in Nature, Nature Methods, Nature Medicine, and Nature Biotechnology papers checked on 2026-05-22."},{"source":"official_project_sources","score":0.97,"rationale":"Operational framing is cross-checked against official Google DeepMind, FutureHouse, and Stanford Biomni project sources."}] |
| allowed-tools | ["read_file","run_shell_command","web_fetch"] |
Scientific Discovery Agents (2026)
Use this skill when the task is not just literature summarization, but a research loop that may generate hypotheses, choose analyses, propose experiments, execute notebooks, or interpret experimental results.
Workflow
- Classify the autonomy level: literature-only copilot, notebook/data-analysis agent, hypothesis generator, experiment planner, lab-in-the-loop agent, or clinical/pathology review assistant.
- Define the scientific claim boundary before execution: hypothesis, in silico result, in vitro result, external validation, or clinical-grade evidence.
- Bind each agent role to explicit tools, datasets, and stopping rules. Do not let a general agent silently become an experimental decision-maker.
- Require human checkpoints at transitions from literature to experiment, from analysis to biological interpretation, and from model output to any clinical or therapeutic claim.
- Evaluate against a specialist baseline: human expert, published analysis, benchmark task, ablation, or wet-lab validation endpoint.
- Preserve a reproducibility package: prompt/config, model, tool versions, data snapshot, generated code/notebooks, citations, failed attempts, and reviewer decisions.
May 2026 Landscape Signals
- Co-Scientist: multi-agent hypothesis generation built with Gemini, with biomedical validations including AML drug repurposing, liver fibrosis targets, and antimicrobial resistance mechanisms.
- Robin: FutureHouse lab-in-the-loop multi-agent system that couples literature search, experimental planning, data analysis, and updated hypotheses for experimental biology.
- CellVoyager: single-cell analysis agent that autonomously proposes and implements scRNA-seq notebook analyses and is evaluated on CellBench.
- SPARK: pathology-agent framework that uses language-mediated agents to generate biologically meaningful tumor-analysis concepts without additional model training.
- Biomni: general-purpose biomedical agent pattern for broad tool ecosystems, useful as an orchestrator only when task boundaries and validation gates are explicit.
Guardrails
- Separate "agent proposed" from "experimentally validated" in every report.
- Do not present AI-discovered therapeutic candidates as clinically actionable without preclinical and clinical validation.
- Treat wet-lab, clinical, and patient-facing outputs as human-reviewed workflows, not autonomous deployment targets.
- For notebook agents, log executed code, environment, data hashes, and failed cells; reject hidden analysis steps.
- For multi-agent systems, track agent roles, messages, tool calls, critique/ranking decisions, and cost.
