You can look at a problem in an agent that uses tools and pick the fix that suits it. Here's a quick look back at the course.
What you can do now
- Get tool calls right. Describe when to call each tool and when not to, test with an eval, and turn on strict tool use.
- Shorten the wait and show what's happening. Run independent tool calls in parallel, and stream so users see which tool Claude chose while they wait.
- Use the tools Anthropic provides. Add code execution and web search, which run on Anthropic's side, and keep computer use for software with no other way in.
- Spend less. Measure the input tokens in each request, cache your tool definitions, and use programmatic tool calling.
Back to your own code
Go back to your own code and start with the problem that bothers you or your users most. Apply the fix from the lesson that deals with it. The lessons stay here, so you can reopen any of them while you work.
Where to look next
| What you need | Where to find it |
|---|---|
| Help with a tool problem the lessons never covered | Troubleshooting tool use(opens in new tab) lists each symptom with its likely cause and the fix. |
| Other tools Anthropic provides, beyond the course's three | Tool reference(opens in new tab) lists them all. |
| Guides and code examples for building with tools | Claude Cookbook(opens in new tab) opens on its Tools category. |
| Connecting Claude to outside services and data without defining and running every tool yourself | Introduction to Model Context Protocol(opens in new tab) is an Academy course where you build MCP servers and clients using the Python SDK. |