Agents in the Docs: Parts, Loop and Limits

People use the word agent loosely. The OpenAI guide gives a clear picture. An agent can plan and finish tasks using tools. It can work with other agents. It keeps track of context across steps. This lesson turns that into a checklist you can use when you read any agent documentation.

The parts of an agent

• Model: the brain. It decides what to do next.
• Tools: search, files, code, your own services and MCP servers.
• Memory: the chat so far, saved sessions and notes. Each product stores these in its own way.
• Handoffs: one agent passes a job to a specialist.
• Guardrails: checks on inputs and outputs, and limits on what it may do.
• Tracing: a record of every step, so you can see what happened.

The loop

An agent repeats a simple cycle. It looks at the goal, picks an action, uses a tool, reads the result and decides again. It stops when the goal is met or a limit is reached.

One agent or several

Start with one agent. It is easier to build and to understand. Add more only when one agent is overloaded. Signs include too many tools, mixed jobs or very long instructions. Then split the work into a few focused agents.

• One agent: a contained job with a few tools.
• Several agents: separate skills, such as research, writing and checking.
• Always give each agent one clear job.

Questions to ask of any agent product

• What can it do on its own, and what needs my approval?
• What does it remember, and where is that stored?
• Can I see each step it took?
• What limits stop it from running forever or spending too much?
• How do I stop it?

A worked example

Say you want an agent that books meetings. The goal is simple: find a time that works and send the invite. The tools are a calendar lookup and an email sender. The memory holds the people and the rules, such as no meetings before nine. The guardrail says to ask a person before it sends an invite to anyone outside the company. The limit says to stop after ten steps.

Now watch the loop. It looks at the goal and checks calendars. It picks a slot, then checks the rules. It asks for approval and sends the invite. If a calendar is full, it tries again with another slot. It stops when the invite is sent or the limit is reached.

This one agent is enough. You only need a second agent if the job grows, such as adding travel booking with its own tools.

When you read about any agent product, map it onto these parts. Ask which part is strong and which is missing. A product with great tools but no step limits is risky. A product with limits but no tracing is hard to trust.