AI Transformation · Glossary

AI Agent

Updated October 25, 2026

An AI agent is an artificial intelligence (AI) model using tools in a loop: it decides its next step from what its tools return, acts, checks the result and repeats until it reaches a stopping condition. In the framing Barry Zhang of Anthropic, a company that sells the computing power agents run on, gave in a conference talk, an agent needs three things: an environment to work in, tools to act with and get feedback from, and a system prompt carrying its goals, constraints and ideal behavior. One of those tools can be a search of company documents, the basis of retrieval-augmented generation (RAG).

An agent decides its own path; a workflow follows a path someone wrote. In Zhang's account, that freedom is the cost: because the agent chooses, you cannot price a run in advance, list what it will do or bound the damage by inspection. AI automation built as a fixed sequence with a model call inside is a workflow, not an agent. An agentic workflow is the pattern of work an agent runs, in which, as the What is an agentic workflow? article puts it, the model "chooses the next step and the tool for it, then loops until a stopping condition is met".

When is an agent worth its cost?

When the path cannot be written down in advance and something can check the result. Zhang's rule: "If you can map the decision tree, build the decision tree." He calls coding the natural agent domain because tests check the output, so for any other use, ask what plays the role tests play. A second question is how long the agent runs before a person is needed again, since that sets how long it can be wrong before anyone notices. Worth knowing who says it: Zhang's company sells that computing power, and he still argued in public that most teams should build workflows.

In one practitioner's framing from MIT Professional Education coursework (2025), early agent builds fail on four things before the model matters:

1. Picking the right tool.
2. Passing one tool's output cleanly into the next.
3. Managing memory inside a finite context, the limited text a model can work with at once.
4. Staying on task, with an evaluation that says whether it finished.

Treat an agent as a tool with a named owner, never an "employee" or "teammate": in one 2026 Boston University experiment, as MIT Technology Review reported it, that framing led people to catch fewer errors, an effect a second account says was concentrated at companies that already list AI on the org chart. Where errors are high-stakes and hard to detect, which Zhang says rules out full autonomy, keep a human in the loop, a person who checks output against a reference before it takes effect. In an AI transformation, start with a workflow and add an agent only where the path cannot be mapped.