Skip to content

Agents & computer use

The frontier of working with Claude is agents: instead of answering a single question, Claude pursues a goal by taking multiple steps, using tools, reading results, and adjusting as it goes.

From assistant to agent

AssistantAgent
You ask, it answersYou set a goal, it acts
One turnMany steps, self-directed
You execute the stepsIt executes the steps

An agent loops: decide the next action → take it → observe the result → repeat, until the goal is met.

What powers an agent

A capable agent combines the pieces from earlier chapters:

  1. The model: the reasoning core that decides what to do next. (Models.)
  2. A runtime: an environment where it can run code and use a filesystem.
  3. MCP connections: access to external tools and data.
  4. Skills: the procedural knowledge for doing tasks well.

This is exactly why the higher Claude tiers (Opus, Mythos-class) matter for long-horizon agentic work: sustained, multi-step tasks need strong, reliable reasoning.

Computer use

Claude can also operate a computer interface, viewing the screen and controlling a mouse and keyboard, to perform tasks in software that has no API. It's powerful for automating GUI-based work, though it's slower and should be sandboxed and supervised.

Examples of agentic work

  • An overnight coding agent that implements and tests a feature.
  • A research agent that gathers sources across the web and compiles a report.
  • A data pipeline that pulls, cleans, analyzes, and summarizes on a schedule.

How to think about it as a non-developer

You don't need to build agents to benefit from the mindset:

  • Delegate outcomes, not keystrokes. Describe the end state you want.
  • Provide guardrails. Constraints, definitions of done, and review points.
  • Keep a human in the loop for anything consequential.

Autonomy needs supervision

The more an agent can do, the more carefully you should scope its permissions and review its actions. Grant least privilege, sandbox where possible, and verify results. See Limitations & trust and Privacy & data.

That completes the toolkit. Finish with the habits that make it all work: Best practices.

Educational material about Claude. Not affiliated with Anthropic. Always verify against official docs.