What is an AI agent?

An agent is a model in a loop with tools, deciding its own next step. That one sentence covers everything from a coding assistant fixing a failing test to a research tool reading fifty web pages.

How does an AI agent work?

An AI agent runs a simple loop around a large language model that can make tool calls:

1. 1
   Receive a goal

   For example, "Find why the checkout test fails and fix it."
2. 2
   Decide an action

   The model reads the goal and everything so far, and picks a tool call such as running the tests.
3. 3
   Act

   The application executes the tool and captures the result.
4. 4
   Observe

   The result, such as a failing stack trace, is added to the model's context.
5. 5
   Repeat or stop

   The model chooses the next action, or decides the task is done, or asks the user for help.

This pattern was described in the 2022 ReAct paper, which interleaved reasoning with actions such as searches. What changed since then is reliability: models are now trained specifically to plan, use tools and recover from errors over many steps.

What is the difference between an agent and a workflow?

In a workflow, your code decides the sequence of steps and the model fills in each one. In an agent, the model decides the steps. Anthropic's widely cited 2024 guide, "Building effective agents", draws this line and recommends starting with the simplest option that works.

Workflows are more predictable, cheaper and easier to test. Agents handle open-ended tasks where you cannot know the steps in advance, at the cost of more tokens, more latency and more ways to go wrong.

When should you build an agent?

Build an agent when the task is open-ended, needs several steps that depend on what earlier steps find, and has a way to check progress. Coding is the standout example because tests and compilers give clear feedback. Avoid agents where a fixed workflow would do, where mistakes are costly and hard to detect, or where the task needs one quick answer.

• ✓Do: Give the agent a small set of well-described tools and clear success criteria
• ✓Do: Let it check its own work: run tests, re-read the output, verify results
• ✓Do: Set limits on steps, time and spending, and a way to ask a human
• ✓Do: Log every step so you can see why it did what it did
• ✕Avoid: Start with a fully autonomous agent when a two-step workflow would work
• ✕Avoid: Give an agent broad write permissions without confirmation on risky actions
• ✕Avoid: Judge an agent from a few impressive demos rather than measured success rates

Why do AI agents fail?

Agents fail mostly through compounding errors. A small mistake in step 3, such as misreading a file, becomes the premise for steps 4 to 20. Other common causes are ambiguous tool descriptions, tool results that flood the context window, losing track of the goal in a long history, and stopping too early or never stopping. Good context engineering and focused tools prevent many of these, and evals that run the whole agent on realistic tasks are the only reliable way to measure progress.

What are multi-agent systems?

A multi-agent system splits work among several model instances, typically a lead agent that plans and delegates to subagents, each with its own context. Anthropic reported in 2025 that its multi-agent research system outperformed a single-agent setup by 90.2% on an internal research evaluation, but used about 15 times more tokens than a chat interaction. Multi-agent designs pay off for broad tasks that split into independent parts, and are wasteful for tasks that need one tight chain of reasoning.