Explainer · Published 26 Oct 2026
AI agent vs chatbot: who decides the next step?
A chatbot produces text for you to act on. An AI agent acts itself: it chooses tools, calls them, reads the results and decides what to do next, in a loop, until it decides it is finished.
| Question | Chatbot | Workflow | Agent |
|---|---|---|---|
| Who decides the next step? | You | The developer, in code | The model, at run time |
| Touches other systems? | Usually not | Yes, along fixed paths | Yes, along paths it picks |
| Typical failure | A wrong answer you might notice | A step that doesn’t fit an unusual case | A plausible wrong action, possibly irreversible |
| Where to review | The output | The design, then spot-checks | The permissions, the gates and the trace |
| Cost profile | One call per turn | Predictable | Variable; loops can run long |
Where the line comes from
Sourced Anthropic’s engineering guide “Building effective agents” (19 Dec 2024) groups both under “agentic systems” but separates workflows, where models and tools follow predefined code paths, from agents, where the model directs its own process and tool use. It recommends starting with the simplest thing that works and adding autonomy only when it pays for itself.
Agentic AI examples (all fictional, all from the lab)
- A refund desk that looks up an invoice, checks policy, issues a refund and emails the customer — four tools, two irreversible actions.
- A travel booker that compares fares on a web page, pays, and updates a calendar — the web page is untrusted input.
- A repository janitor that lists stale branches and deletes them — deletion is the irreversible step.
Interpretation In each case the useful question is not “is the model smart enough?” but “what is it allowed to do, and who looks before it does the thing that can’t be undone?”
A simple AI agent workflow, step by step
- Goal in. A person states the task and constraints.
- Plan. The model proposes steps (or a developer fixed them, in a workflow).
- Act. Each step calls a tool within a permission scope and a budget.
- Observe. Tool results return into the model’s limited context.
- Gate. Irreversible steps pause for approval — ideally enforced by the harness, not left to the model’s memory.
- Stop & record. The run ends and leaves a trace someone can audit.
The Delegation Lab lets you run that loop deterministically and see where it breaks.