An AI agent is a system that pursues a goal across multiple steps: it plans, uses tools, checks results, and keeps going until it finishes or decides to escalate. That's the meaningful distinction from a chatbot — a chatbot answers; an agent acts.

In practice, "agent" usually means a loop around an LLM that can call tools via function calling (or a protocol like MCP), observe the outcome, and decide what to do next. A prompt chain with hardcoded steps is useful, but it's not the same thing. An agent reasons about its next action at each step.

Agents change your risk profile significantly. A bad answer from a chatbot is annoying. A bad action from an agent can move money, delete records, or send emails to the wrong people. The lethal trifecta — cost, latency, and reliability — compounds across every step in a long task. A system that's 95% reliable per step is 60% reliable across ten steps. That math matters when you're deciding what to hand off to autonomous execution.

This means guardrails aren't optional once you introduce tool use. It also means the scope of what you hand to an agent should scale with how confident you are in your evals and your fallback handling. Start narrow, instrument everything, and expand from there.

Further reading:

• Andrew Ng on AI Agentic Workflows (March 2024) — Introduced the four agentic design patterns: reflection, tool use, planning, and multi-agent collaboration.
• Building Effective Agents (Anthropic) — Research on composable patterns for building reliable agent systems.
• What are AI Agents? (Google Cloud) — Overview covering autonomous systems that use LLMs for reasoning and action.