Agentic Representation Learning: How Should AI Agents Be Represented and Learned from Black-Box Agents?

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

AI agents are becoming persistent, interactive systems that plan, remember, use tools, adapt, and coordinate, yet we still describe them mainly by implementation details and benchmark scores. We argue that agent research is missing reusable representations of the agent itself. We frame Agentic Representation Learning (ARL) around two questions: what representations should an agent have and how can such representations be learned when the agent is accessible only as a black box? We place structural components, representational factors, external signals, and downstream tasks within a unified framework, and unify the learning processes for both white-box and black-box agents. Meanwhile, we introduce a characterizer agent that selects diagnostic interactions to learn about the black-box agent, and we present a concrete workflow for acquiring representations through iterative probing, observation, and representation updating. The ARL framework provides a foundation for enabling AI agents to understand each other effectively and for developing multi-agent systems.