How AI agents work, in plain English
AI agents, explained plainly
An AI agent is a program that uses a language model to take steps toward a goal. These pages explain what that means, how the loop and tool calling work, and what limits apply. Everything comes from the makers' own documentation.
All the notes
- What is an AI agent? A plain-English explanation
What the makers mean by "AI agent", how it differs from a plain chat window, and why the word has no single agreed definition.
- How an agent works: the plan, act and check loop
What happens between your request and an agent's answer, as the makers describe it, covering planning, tool calls, checking results and stopping rules.
- Tool calling explained: how an AI asks other software to do things
How a language model that only produces text can use a calendar, a search or a database, according to Anthropic's and OpenAI's own documentation.
- Agent or fixed workflow: what is the difference?
How Anthropic separates workflows (fixed paths) from agents (model-directed steps), with the named workflow patterns and a few questions to ask of a product.
- What an agent is allowed to do: permissions, access and approvals
How makers let you limit what an agent can touch, using permission rules, approval prompts and careful connections, with the cautions from their documentation.
- Why agents go wrong: mistakes, loops and misread instructions
The kinds of error the makers say agents make, such as compounding slips, guessed inputs and tricked instructions, and the checks they suggest.
- Agent words glossary: tools, memory, context, orchestration and more
Plain-English meanings for the terms in makers' agent documents, from agentic system and tool call to context window, MCP, guardrails and prompt injection.