Agents, workflows and chat
Anthropic's engineering guide draws a useful line between two kinds of AI systems:
- A workflow runs a language model and tools along steps a developer fixed in advance: summarise, then classify, then file a ticket.
- An agent lets the model choose the steps. It reads the task, picks a tool, looks at what came back, and decides what to do next, in a loop, until it is done or it needs you.
A plain chat with an assistant is neither: you ask, it answers. When the same assistant starts searching the web, running code or calling tools on its own to finish your request, it is acting as an agent.
What makes an agent work
- Tools: without them an agent can only talk. MCP servers and built-in tools give it actions to take.
- Feedback: the result of every step (a search result, an error, a test run) goes back to the model, so it can correct course.
- A stopping point: a limit on steps, a finished task, or a checkpoint where it asks you.
When to use one
The same guide recommends starting with the simplest thing that works. A single well-written prompt, or a fixed workflow, is cheaper, faster and more predictable. Agents suit open-ended tasks where you can't list the steps ahead of time. They cost more and their mistakes can add up over many steps, so give them checkpoints where a person reviews the work, and try them somewhere safe first.
In the directory
The AI agents section lists hosted agents, open-source agents and frameworks for building them. Many agents can also use agent skills and MCP servers.