A standard LLM takes input, generates output, and stops. An AI agent takes a goal and figures out the steps to achieve it — using tools, making decisions, and looping until the job is done.
The core loop of an agent:
1. **Observe**: Receive a goal or current state
2. **Think**: Reason about what to do next (often using chain-of-thought)
3. **Act**: Call a tool — run code, search the web, write a file, call an API
4. **Observe result**: See what happened
5. **Repeat** until the goal is complete
This is called the **ReAct pattern** (Reasoning + Acting). It's how systems like AutoGPT, Devin (the AI software engineer), and Claude's computer use work.
Tools are what give agents power. An agent with a code execution tool can write a script and run it. With a web browser, it can research and act on real-world information. With a calendar API, it can actually schedule your meeting.
The hard problems: agents fail silently, can get stuck in loops, and compound errors badly over long tasks. A wrong step early can cascade into a completely wrong outcome. This is why human-in-the-loop checkpoints matter for any real-world deployment.
The most capable agents today use **multi-agent systems** — a supervisor agent breaks a task into subtasks, specialized sub-agents handle each piece, and results are assembled. This mirrors how engineering teams work.
**Key takeaway:** Agents = LLMs + tools + a loop. They don't just answer questions — they take action in the world.