What makes an agent
The difference between a chatbot and an autonomous AI worker
Beyond question-and-answer
A chatbot waits for you to ask something and gives you one answer. An agent has a goal and takes a sequence of actions to accomplish it — reading files, calling APIs, making decisions, correcting itself when something goes wrong, and reporting back when done.
The difference isn’t the underlying model. It’s the loop: observe → plan → act → observe the result → plan the next action → repeat.
The agent loop
| Step | What happens | Example |
|---|---|---|
| Observe | The agent sees the current state | Reads your unread emails |
| Plan | Decides what action to take next | “These 3 need replies, 2 are spam, 1 needs research” |
| Act | Takes an action using a tool | Drafts replies, archives spam |
| Observe result | Sees what the action produced | Confirms drafts were created |
| Next action | Plans based on new state | Flags the research email for you |
What gives an agent capabilities: tools
- Search / retrieve — read emails, query databases, search the web
- Write / create — draft documents, write code, create files
- Communicate — send emails, post to Slack, create calendar events
- Execute — run code, trigger workflows, call APIs
- Navigate — browse websites, fill forms, click buttons
Agents vs. automation scripts
| Traditional automation | AI agent | |
|---|---|---|
| Handles unexpected input | Fails or errors | Adapts and continues |
| Requires exact format | Yes | No — handles variation |
| Can make judgment calls | No | Yes (within limits) |
| Needs human oversight | Low (deterministic) | Medium (probabilistic) |
Key insight: Agents are powerful because they handle the variation and judgment calls that break traditional automation. But they’re probabilistic, not deterministic — which means they need appropriate human oversight for high-stakes tasks.