A structured framework for understanding AI agent capabilities, maturity, and deployment pathways, sector by sector.
When Dmitri Mendeleev organized the periodic table in 1869, he wasn’t cataloguing novelties. He was imposing structure on chaos. By arranging known elements by atomic weight and recurring properties, he made invisible relationships visible. Gaps in the table predicted elements not yet discovered.
This project applies the same logic to AI agents. The capabilities exist. The deployments are underway. What’s missing is the map: a common grammar for understanding what kind of work an agent does, how autonomously it operates, and where it fits in an institutional workflow. These tables are that map, built sector by sector.
What work is being done: the functional area or capability type the agent operates within.
How autonomously the agent performs, from answering questions to orchestrating multi-step workflows.
As capabilities mature, AI evolves from a helper into a system that executes work across teams and systems.
Cross-cutting agent types that operate across multiple domains: reasoning, synthesis, orchestration, without belonging to a single functional row.
Experimental or high-risk agent configurations. High capability, but requiring greater governance, oversight, and institutional care before deployment.
Identify the workflow or outcome you want to improve before selecting an agent type.
Find the domain column that matches your work. Read up the maturity axis to see where you are and where you could go.
Sequence matters more than scale. A well-deployed simple agent outperforms a poorly governed orchestrated one every time.
The Lanthanide and Actinide agents are often where the highest leverage and highest risk lives. Approach them deliberately.
A note on scope. These tables are frameworks, not forecasts. They reflect current deployment patterns and emerging capabilities as of mid-2025. Agent maturity varies significantly by institution, vendor, and use case. The goal is orientation, not prescription. Treat the table as a starting map, not a finished one.