Part 4: Agents and computer use

The last tool group shows the pattern behind several of the previous tools. Agents are chat applications with tools. They can take actions on your behalf.

For coding assistants, those actions can include:

• listing files
• reading files
• writing files
• searching a codebase
• running shell commands

Product names change faster than the pattern.

Once you recognize the agent loop, compare tools by three questions:

• What actions can they take?
• What context can they see?
• What guardrails do they give you?

Agents in coding tools

In the public README, we put it in practical terms:

• Coding assistants and project bootstrappers both use agents for coding
• Agents have tools such as reading files, writing files, and indexing the codebase
• The model chooses when to use those tools

The Snake examples make the difference visible. A chat app returns a component, while a coding agent reads the existing src/ folder and the files in it. The agent can edit App.jsx, create components/SnakeGame.jsx, and run a build command.

That doesn't remove your responsibility, but it changes where you spend time. You spend less time copying code and more time checking diffs. You also read the tool output and decide whether the direction still matches the project.

In the earlier demos, the same pattern shows up several times:

• Claude Code lists files, runs npm commands, and edits Vite files
• Copilot reads the project and refactors App.jsx
• Cursor looks at the score state, grid rendering, and border behavior
• Lovable edits a hosted template and publishes a preview

To learn more about agents

In the public README, we point to two learning resources:

• RAG agents workshop
• LLM Zoomcamp

Use those if you want to understand function calling and tool use directly instead of only through a coding product.

Anthropic computer use demo

You can also try Anthropic's computer use demo. It gives the model a browser-like computer environment instead of only project files.

Run it with Docker and an Anthropic API key:

docker run \
    -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY \
    -v $HOME/.anthropic:/home/computeruse/.anthropic \
    -p 5900:5900 \
    -p 8501:8501 \
    -p 6080:6080 \
    -p 8080:8080 \
    -it ghcr.io/anthropics/anthropic-quickstarts:computer-use-demo-latest

The ports expose the demo interfaces from the container, and the exact UI can change over time. Check the upstream quickstart if the container output prints a different local address.

After the container starts, Docker runs a Linux desktop environment. An agent inside that desktop can interact with the UI, and you get a chat interface next to a desktop view.

Start by taking screen control and opening Firefox manually, so the agent doesn't have to handle browser setup.

Then release control and ask:

Find me a picture of a cute kitten.

The agent takes a screenshot, decides it needs a search, types a search query, and opens image results. It's slower than a human, but it shows the important capability: the model can operate a GUI when the tool exposes screenshots and input actions.

PR Agent

In the public README we also list PR Agent as another agent example. That tool works on pull request review and description rather than app generation.

The same idea applies outside coding. If the tools are spreadsheet actions, the agent can work on a spreadsheet. If the tools are GitHub actions, the agent can work on pull requests and issues. The tool set defines what the agent can do.

This closes the comparison loop. Chat apps generate code from conversation, IDE assistants edit a local project, and bootstrappers create hosted prototypes. Agents describe the tool-using pattern behind all of them.