Agentic Science represents a new research paradigm where AI manages complete investigation cycles autonomously
Core Insight
“Agentic Science” is the emerging paradigm where AI systems are not merely tools assisting human researchers but active participants that plan, execute, and interpret complete scientific cycles — from hypothesis generation to conclusion — in closed loops with minimal human intervention. This parallels PUMA’s vision for the Smart PMO: not just automating individual PM tasks but orchestrating complete sprint planning cycles.
Key Agentic Science Systems
| System | Domain | Level of Autonomy | Key Capability |
|---|---|---|---|
| AI Scientist v2 (Lu et al., 2026) | ML Research | High | Paper accepted at ICLR workshop with zero human writing |
| Adam (King et al., 2009) | Systems Biology | High | Discovered yeast gene functions autonomously |
| Eve (Williams et al., 2015) | Drug Discovery | High | Identified drug candidates via active learning |
| GNoME (DeepMind, 2023) | Materials Science | High | 2.2M new stable structures, hundreds verified |
| ChemNavigator (Zheng et al., 2026) | Photocatalyst Design | Medium | Derived interpretable design rules for H2 evolution |
| Genesis (Tiwari et al., 2024) | Systems Biology | Emerging | Full systems biology area via microfluidics + ML |
The Closed-Loop Science Model
[Hypothesis Generation] → [Experiment Design] → [Execution] → [Analysis] → [Conclusion] → [New Hypothesis]
↑___________________________ AI AGENT LOOP __________________________________|
(with human oversight at key gates)
This mirrors PUMA’s Stage 5 Smart PMO:
[Sprint Start] → [Issue Triage] → [Estimation] → [Risk Detection] → [Sprint Plan] → [Retrospective]
↑___________________________ SMART PMO AGENT LOOP __________________________|
(PM approves sprint plan before execution)
Limitations of Agentic Science (2024-2026)
- Framework dependence: Search space, evaluation criteria, and knowledge base still defined by humans
- Interpretability gap: Emergent results lack causal explanation
- Reproducibility challenges: Stochastic LLMs produce variable results without seeding
- Hallucination risk: Fabricated references and unverifiable intermediate steps
- Attribution problem: Who “owns” AI-generated discoveries?
PUMA Implications
PUMA occupies the intersection between agentic science methodology and project management practice. The same architectural patterns (closed-loop hypothesis-experiment-analysis) appear in both:
- Scientific discovery: agent → experiment → metric → revision
- PM automation: agent → triage → sprint plan → velocity feedback
PUMA Stage 5 is, in this sense, an applied Agentic Science system for the PM domain.
Related Notes
- PN-AI-Scientific-Knowledge-Generation — evidence base for AI knowledge generation
- PN-PUMA-within-AgenticScience-Trajectory
- PN-MultiAgent-ArchitecturePatterns — Specialised agents in science
- PN-ReAct-AgentPattern — reasoning loop as core agentic mechanism
- LN-Zhang-2025-AgenticScienceSurvey
- LN-Lu-2024-AIScientist
- LN-Jumper-2021-AlphaFold
- LN-Jain-2023-GNoME-Materials
- LN-Klinger-2025-AIScience-SocialProblem
- Smart-PMO-Vision — PUMA Stage 5 as applied Agentic Science
- PR-PUMA-Ch1-Introduction — Three enabling conditions
- EX-Hypotheses-H1-H2 — PUMA’s experimental grounding