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

SystemDomainLevel of AutonomyKey Capability
AI Scientist v2 (Lu et al., 2026)ML ResearchHighPaper accepted at ICLR workshop with zero human writing
Adam (King et al., 2009)Systems BiologyHighDiscovered yeast gene functions autonomously
Eve (Williams et al., 2015)Drug DiscoveryHighIdentified drug candidates via active learning
GNoME (DeepMind, 2023)Materials ScienceHigh2.2M new stable structures, hundreds verified
ChemNavigator (Zheng et al., 2026)Photocatalyst DesignMediumDerived interpretable design rules for H2 evolution
Genesis (Tiwari et al., 2024)Systems BiologyEmergingFull 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)

  1. Framework dependence: Search space, evaluation criteria, and knowledge base still defined by humans
  2. Interpretability gap: Emergent results lack causal explanation
  3. Reproducibility challenges: Stochastic LLMs produce variable results without seeding
  4. Hallucination risk: Fabricated references and unverifiable intermediate steps
  5. 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.


MOCs