What Is an AI Agent? – Definition and How It Works
Learn what an AI agent is, how the perceive–decide–act loop works, how agents differ from chatbots and workflows, and where they fail in production.
An AI agent is a system that uses a language model to decide what to do next, takes an action, observes the result, and repeats until a goal is reached. The defining feature is not intelligence — it is the loop.
1. What Is an AI Agent?
An agent pursues a goal by choosing its own sequence of actions rather than following a fixed script.
- Key idea: the model decides the next step, informed by what the previous step returned.
- Mechanism: a loop of observe → decide → act, with tools as the means of acting.
- Goal: complete a task whose required steps are not known in advance.
The contrast that clarifies it: a workflow knows its steps at design time, an agent discovers them at run time.
2. The Loop
- Goal. The agent receives an objective and any context.
- Decide. The model reasons about the current state and selects a tool with arguments.
- Act. The application executes the call — a search, a query, a page fetch, a file write.
- Observe. The result is appended to context.
- Repeat until the model judges the goal met, or a limit stops it.
Everything distinctive about agents follows from step 4 feeding back into step 2. Without that, you have a single tool call, not an agent.
3. The Components
| Component | Role |
|---|---|
| Model | Chooses the next action |
| Tools | Everything the agent can actually do |
| Memory | Context within a run, and sometimes across runs |
| Planning | Decomposing a goal into steps, explicit or emergent |
| Loop control | Step caps, budgets, timeouts, stop conditions |
The tools are the real boundary. An agent with no tools is a chatbot; an agent's capability set is exactly its tool set, and so is its blast radius.
4. Agent vs. Chatbot vs. Workflow
| Chatbot | Workflow | Agent | |
|---|---|---|---|
| Steps | One response | Fixed, predefined | Chosen at run time |
| Acts on the world | No | Yes, prescribed | Yes, self-directed |
| Adapts to results | No | Only via branches | Yes |
| Predictable cost | Yes | Yes | No |
| Debuggable | Trivially | Easily | Requires tracing |
| Right when | Answering | Steps are known | Steps are unknown |
The most common and most expensive engineering mistake is reaching for an agent when the steps are known. If you can draw the flowchart, build the flowchart — it will be cheaper, faster, and testable.
5. Where Agents Genuinely Help
- Variable-depth research. How many sources are needed is not knowable up front.
- Navigating unfamiliar interfaces, where the next click depends on what the last one produced.
- Error recovery, where the response to a failure depends on the failure — retry, alternative route, or stop.
- Heterogeneous sources, where each one needs a different approach discovered on inspection.
- Long-tail tasks too varied to justify a workflow each.
6. Where They Fail
Cost is unbounded by default. Every step is a model call carrying the whole accumulated context. A ten-step run is not ten times a one-step run — context grows each iteration.
Loops. An agent that cannot make progress will often keep trying variations of the same failing action. Step caps are not optional.
Silent wrongness. An agent that fails a subtask and continues produces a confident final answer built on a gap. This is far more dangerous than a crash, and it is the failure mode most production incidents trace to.
Non-determinism. The same input can produce different action sequences, which makes testing statistical rather than exact.
Compounding error. A 95% reliable step is 60% reliable ten times over. Long agent chains need either fewer steps or verification between them.
7. Real-World Examples
- A research agent that searches, reads results, notices a gap, and searches again with better terms — depth decided by findings.
- A browser agent navigating a site with no API, deciding each interaction from the rendered page.
- A coding agent that runs tests, reads failures, edits, and re-runs — a loop whose length depends on the bug.
- A data collection agent that inspects a page, chooses between an internal JSON endpoint and DOM parsing, and adapts when a layout changes.
- A support agent that looks up an account, checks orders, and escalates — often better as a workflow, and a good example of over-reach.
8. Summary
An AI agent uses a model to select actions, executes them through tools, observes results, and iterates toward a goal. The loop is the definition; the tools are the capability and the risk.
Agents earn their cost when the path cannot be known in advance — research, navigation, recovery, heterogeneous sources. When the steps are known, a workflow beats an agent on cost, speed and testability every time.
Production agents need step caps, budget limits, tracing, and explicit handling for partial failure. The characteristic production incident is not a crash but an agent that quietly failed one step and reported a confident answer anyway.