THE SHORT ANSWER

In this guide, an AI agent is a system that uses a model to choose actions towards a goal, observes their results and decides what to do next. Definitions vary; the useful questions are which decisions the system makes and which permissions it has.

A working definition, not a universal label

Anthropic distinguishes workflows with predefined paths from agents whose models direct the sequence of actions and tool use. We use that distinction here. Other products use ‘agent’ more broadly, so the label alone does not establish autonomy or capability.

A system can be highly automated without being model-directed. A scheduled process that always performs the same three steps may be a good workflow. Adding an agent is useful only when choosing the next step is part of the problem.

Evidence & context: Anthropic

Goal → action → observation → next decision

Imagine a research assistant tasked with finding a school's publicly available term dates. It could search, open an official page, notice that it describes a previous year, and look for the current calendar. Each observation changes what it does next. This is an illustrative design, not a claim that every agent behaves reliably.

Now separate that process from authority. Finding a calendar is one permission. Editing a school's website is another. A model's ability to suggest an action does not mean it should be allowed to execute it.

  • Goal: the outcome and conditions for success.
  • Tools: the operations actually available.
  • Feedback: what the system observes after each step.
  • Boundaries: actions requiring approval and conditions for stopping.

One wrong assumption can become several wrong actions

If the assistant assumes the wrong school is the right one, every later step can be competent and still serve the wrong goal. Testing only whether individual tools run successfully will miss that failure.

A better test includes ambiguous names, missing documents, conflicting dates and tool errors. The expected result may be a request for clarification or a refusal to continue—not completion at any cost.

When is an agent worth considering?

Use the agent-versus-chatbot comparison to separate the interface from the execution model. If the goal is simply to explain a document, a constrained assistant may be enough. If the task needs decisions between uncertain steps, an agent may offer flexibility, with additional testing and oversight.

Begin with a small, reversible task and an explicit stopping condition. A clear boundary is more valuable than an impressive list of tools.

Sources & further reading

  1. Building effective agents

    Anthropic. A provider's engineering taxonomy of agents and workflows, not a universal industry definition. We use the conceptual distinction, not its changing product recommendations.

Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.

A correction, a counterexample or an experience worth sharing?

Join the conversation ↗