AI agents
What is an AI agent made of?
Every agent, however it is built, comes down to three parts: a model that does the reasoning, tools that let it act, and instructions that hold the goal and the rules.
Every agent, however it is built, comes down to three parts.
- A model that does the reasoning. This is the brain that decides what to do next.
- Tools that let it act: read a file, run a search, call an API, send a message. Without tools an agent can only talk. With them it can change things. See tool use and MCP.
- Instructions that hold the goal and the rules: what success looks like and what the agent may or may not do.
OpenAI’s guide frames these as the model, the tools, and the instructions that give the agent a reliable plan.[1] Strip out the tools and you are back to a chatbot. Strip out the instructions and the model has nothing to aim at.
Each ingredient fails differently, which is a useful way to debug an agent that is not working. Take a support agent that handles refund requests. The model decides whether a given request looks legitimate or needs a human — a weak model reasons about that badly no matter how good its tools or instructions are. The tools cap what it can physically do: a lookup for the order and an API call that issues the refund, nothing more; give it the refund tool without the lookup tool and it has to guess at facts it should be checking. The instructions set the boundary: refund automatically up to some amount, escalate anything larger, always say why in the reply.
Loose instructions are the most common failure of the three in practice. A capable model with no stated boundary does not sit idle — it improvises one, and an improvised boundary is rarely the one you wanted. That is the subject of rules and guardrails: instructions written so they actually hold, rather than reading as a suggestion the model can talk itself out of.