THE SHORT ANSWER
A chatbot presents a conversational interface. An AI agent, in the working definition used here, chooses actions and responds to their results to pursue a goal. Some chatbots are interfaces to agents; others answer questions or follow fixed flows without that kind of control.
Compare behaviour, not the label on the box
There is no single agreed boundary around the word agent. Anthropic's workflow-versus-agent distinction is useful for design, but a vendor may describe the same product differently. Ask for a demonstration of who chooses the next step.
| Question | Answering assistant | Fixed workflow | Model-directed agent |
|---|---|---|---|
| What is the central job? | Respond to a request | Complete predefined steps | Choose steps towards a goal |
| Who determines the path? | The user drives turns | The designer specifies the flow | The model selects actions within constraints |
| Can it use tools? | Possibly | Yes | Usually, in the systems discussed here |
| What should be tested? | Answer quality and evidence | Branches, inputs and failures | Goal interpretation, action choices and stopping |
| What permission is appropriate? | Only what the task needs | Only what the steps need | Only what the bounded goal needs |
Evidence & context: Anthropic
The same library request, three designs
Imagine a learner asks, ‘Can I keep this book another week?’ An answering assistant explains the published renewal policy. A fixed workflow looks up the loan, checks a predefined eligibility rule and asks the learner to confirm renewal. An agent might investigate an ambiguous request, consult available records and choose which clarification or tool to use next.
These are illustrative designs. None is automatically more accurate or more appropriate. The fixed workflow may be the best choice when the policy and steps are clear. The conversational assistant may be enough when no account access is needed.
Action changes the consequence of an error
A wrong explanation can mislead. A wrong action can also change a record or send a message. Do not assume a friendly chat interface means a system is only giving advice. Look for the point at which a suggestion becomes an operation.
- Can it only read, or can it write and send?
- Which actions require my approval?
- Can I see what it did and why?
- What happens if it cannot finish, or I ask it to stop?
Choose the least complicated system that meets the need
Before building an AI agent, write a normal successful case and a difficult one. If both can be handled with clear rules, a workflow may be easier to inspect. If discovering the next step is essential, model-directed behaviour may be worth evaluating.
Treat autonomy as a design choice with costs and responsibilities, not a ranking in which more is always better.
Sources & further reading
- 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 ↗