What are autonomous agents?
The 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.
Autonomous agents are systems that can pursue a goal with some independence over their next actions. Instead of receiving a separate instruction for every step, an agent observes what happens, selects an action and continues within the authority it has been given.
Autonomy comes in degrees. A research agent may choose which documents to read but have no permission to contact anyone. A scheduling agent may identify available times but need approval before booking. Describing these boundaries is more useful than calling either system “fully autonomous.”
Independence also differs from competence. A system can operate without interruption and still make poor decisions. Conversely, a highly capable model may be deliberately limited to preparing drafts. The level of supervision is a design choice, not a direct measure of intelligence.
The surrounding software supplies tools, preserves progress and controls what actions can run. A tool is a callable function such as searching records or updating a task. The model can propose an action, but software should enforce permissions and validate important details. Instructions alone are not an access-control system.
Long tasks introduce several opportunities for failure. A commonly used illustration multiplies the success probabilities of individual steps: twenty independent steps that each succeed 95% of the time have about a 36% chance of all succeeding. The independence assumption and the definition of success matter. Real agents can detect errors, retry or make correlated mistakes, so this is an illustration—not a measured law of agent reliability.
In a referral workflow, an agent might check whether required documents are present, identify an unresolved request and prepare a follow-up. Whether it may send that message should depend on the content, recipient, consequences and approved process. Some narrow actions can have standing authorization; others need review. The essential questions are what the system can do, how its actions are checked and who takes responsibility when it cannot complete the task.
Sources: Anthropic — Building Effective Agents (2024); Yao and colleagues — ReAct: Synergizing Reasoning and Acting in Language Models (2023); Anthropic — Demystifying Evals for AI Agents (2026)
To learn more, take a look at 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