What is a subagent?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
A subagent is a separate agent instance assigned a bounded part of a larger task. An agent is software that can choose actions toward a goal and respond to the results. The subagent works on its assignment, often with its own working information, and returns a result to a coordinating agent or application.
Suppose a project involves reviewing twenty documents. The coordinating agent might ask separate subagents to examine different groups of documents, then combine their findings. Each subagent needs to know what question to answer, what material it can use and what its report should include. “Help with the project” is not a clear enough assignment.
A separate context means that the subagent may not automatically see everything the coordinator has seen. Context is the information supplied to a model for its current task. This separation can keep irrelevant material out of an assignment, but it also creates a handoff problem: omitted instructions or qualifications may never reach the person—or agent—doing the work.
The subagent may use the same underlying model as the coordinator. A model is the trained component that interprets information and produces responses. Separate roles do not necessarily require separate model families. The distinction lies in the assignment, available information, tools and execution.
A tool is not necessarily a subagent. A calculator performs a defined operation; it does not independently decide how to investigate the broader problem. A subagent has a delegated task and some discretion about completing it. A skill is different again: it is a reusable procedure that either agent may follow.
Specialization should address an observed problem. If one part of an invoice-processing task keeps going wrong, a focused subagent may help. Adding agents simply because a team sounds sophisticated is a poor reason to do it. Delegation adds coordination work, repeated computation and the possibility that a short returned summary loses important evidence.
The class’s invoice example used observed mistakes to motivate specialization. A healthcare adaptation could assign one subagent to extract documented facts from a fictional note and another to locate the applicable passage in an approved coding reference. The coordinator would compare those outputs without allowing either agent to invent missing findings. Each needs a bounded assignment and inspectable evidence. Agreement between them is not independent confirmation if both copied the same mistaken assumption; adding a specialist role is something to evaluate, not a guarantee.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy