What it is
An agent is an LLM-powered program that can choose actions (often via tools like web, code execution, or APIs), keep state, and iterate until it completes a task.
What makes something an agent
- A goal (explicit or implied)
- A loop (plan → act → observe → adjust)
- Tool access (optional but common)
- Memory/state (short-term context and/or long-term storage)
Common pitfalls
- Tool spam (calling tools without improving the outcome)
- Hidden costs (token usage, latency, API limits)
- Brittle prompts (works in demos, fails in production)