What is an AI agent? A definition for product teams
An AI agent is a system in which a language model decides, step by step, which actions to take to reach a goal: it calls tools, reads the results, then chooses the next action until it judges the task done or hits a limit. Unlike a fixed workflow, the sequence of steps is not written in advance. Anthropic, among others, separates workflows, in which code orchestrates the model, from agents, in which the model steers the process itself.
Definition written by Alex Dionisio, Senior Product Manager. Updated November 13, 2026.
Why it matters for a PM
Autonomy has a price: higher cost per task, more latency, less predictable results and a larger blast radius when something goes wrong. The PM decides whether a task really needs that flexibility, which tools the agent may use, when a human must approve an action, and how success will be measured before launch.
Example
A sales operations agent receives “prepare the renewal for Acme”. It looks up the contract in the CRM, pulls usage data, drafts a renewal proposal and stops before sending it, because sending an email to a customer requires approval from the account owner.
How does an AI agent actually work?
An agent runs in a loop. The model receives the goal, the list of available tools with their descriptions, and a record of what has been done so far. It chooses the next action, for instance calling a given tool with given parameters. The application executes that call, returns the output to the model, and the cycle continues until the model judges the goal reached or a limit stops it.
So the model never acts by itself: it proposes actions, and the surrounding code, sometimes called the harness, executes, allows or blocks them. That code is where permissions, human approvals and execution logs live.
Agent, workflow or assistant: what is the difference?
A conversational assistant handles a request, perhaps with one tool, and then hands control back. A workflow chains steps defined in advance, some of which call a model: the path never varies, which makes it predictable and easy to test. An agent decides its own steps and their order based on what it finds along the way.
The right instinct is to start with the simplest option. If you can draw the process on a whiteboard and it barely changes from case to case, a workflow will do the job for less. Many teams start with a workflow and introduce agentic steps only where the fixed path keeps breaking. An agent earns its place when the variety of situations makes that diagram impossible to draw.
How do you evaluate an agent before launch?
Build a set of realistic tasks, each with a verifiable end state: the draft exists in the right folder, the field has been updated, no forbidden action was attempted. Run the agent several times on each task, since two runs can take different paths, and measure success rate, number of steps, duration and cost.
Read full trajectories as well. An agent can reach the right result by a risky route, or fail because a tool is poorly described rather than because of the model. Those reviews point to the right fixes: rewriting a tool description often helps more than switching models.
What should you ask the engineering team?
Which tools can the agent call, with what permissions, and which ones write to a system? Which actions require approval, and in what form? How does the agent react if a tool fails or returns something unexpected? Can a run be replayed to understand a failure? The answers describe the real level of risk far better than the name of the model in use. Ask as well who is alerted when the agent stops halfway, and how a person can pick up the task where it left off.
Key points
- Many use cases are better served by a simple workflow or a single model call; an agent pays off when the path to the goal varies from case to case.
- An agent is only as reliable as its tools: clear descriptions, explicit errors and safe retries.
- Every agent loop needs stop conditions: a maximum number of steps, a budget, a time limit.
- Irreversible or sensitive actions (payments, emails, deletions) call for human approval.
- Agents are evaluated on whole tasks and on their trajectories, not only on the final answer.
Common mistakes
- Calling any chatbot an agent, which blurs the risk discussion.
- Granting broad permissions “to see what it can do” in a production environment.
- Ignoring prompt injection through content the agent reads, such as emails, web pages or tickets.
- Shipping without a trace of each run, which makes failures impossible to diagnose.
Go further with Module
The courses and lessons that cover this concept:
Guides on this topic
- Write a spec for an AI coding agent
The sections of a spec a coding agent can execute without guessing, with a SPEC.md template.
- Claude Skills: build and test a skill
SKILL.md anatomy, install, triggering, testing and security, with a template.
- MCP for Product Managers
Connect AI to your tools and scope an MCP server: tools, permissions, security, with a spec template.