Building with the Claude API and agents

You build agents yourself on the Anthropic API and learn how to connect Claude to real tools, databases and documents. From the first API call to a production agent with evaluation, logging and protection against prompt injection.

Description

The course starts with the basics of the Anthropic API. You learn the Messages API inside out: authentication, choosing between the Opus, Sonnet and Haiku models, streaming, vision, multi-turn conversations and system prompts. You understand how pricing works and how to make serious cost savings with prompt caching.

You then move on to tool use and document grounding. You define functions that Claude can genuinely call, connect it to existing APIs and databases, and build a Q&A bot over your own documents with citations. All in your own IDE, in Python or TypeScript.

In the second part you delve into agent patterns. You implement ReAct loops and plan-and-execute, distribute tasks across sub-agents and write your own MCP server (Model Context Protocol) that exposes your business tools to Claude. You learn how to evaluate agents with eval sets and scoring, and how to set up regression tests.

Finally, you cover everything you need for production: rate limits, fallbacks, error handling, logging, cost control and protection against prompt injection. You finish with a multi-step agent that carries out real tasks in your own context.

Topics

• The Anthropic API: authentication, model selection (Opus, Sonnet, Haiku) and pricing
• The Messages API: system prompts, roles and multi-turn conversations
• Streaming responses and vision (having Claude interpret images)
• Tool use: defining functions and having Claude call real APIs
• Document grounding with citations across your own documents
• Prompt caching to reduce costs with repeated context
• Agent patterns: ReAct loop, plan-and-execute and multi-step reasoning
• Model Context Protocol (MCP): when to use it and writing your own MCP server
• Sub-agents and delegation between specialised Claude instances
• Evaluation: eval sets, scoring and regression tests for your agent
• Production: rate limits, fallbacks, error handling, logging and cost control
• Security: mitigating prompt injection and validating output

Result

After the course, you will work independently with the Anthropic API and build AI applications that you can put straight into production. You know the Messages API inside out and know which model suits which situation. You have written agents using proven patterns and know how to measure their quality.

Specifically, by the end you will be able to:

• Building working integrations on the Anthropic API in Python or TypeScript
• Using tool use to connect Claude to APIs, databases and internal systems
• Applying document grounding and prompt caching for accurate and cost-efficient solutions
• Writing your own MCP server that exposes your business tools
• Designing agents using ReAct, plan-and-execute and sub-agent patterns
• Measuring the quality of agents with eval sets, scoring and regression tests
• Tackling production concerns: rate limits, fallbacks, logging and cost control
• Recognising and mitigating prompt injection

Target audience

This course is intended for backend developers, AI engineers, ML engineers and tech leads who want to integrate AI functionality into their own applications. You work with code on a daily basis and are comfortable with REST APIs, JSON and async patterns.

You have at least two years' experience with backend development in Python, TypeScript or a comparable language. Prior knowledge of AI or LLMs is not required — you learn the relevant concepts around Claude, tool use and agents during the course.

The course is technical and code-heavy. You write working agents yourself and go home with code that you can deploy directly in your own stack. Would you rather approach AI at a strategic level, or mainly learn to code with AI assistance? In that case, AI strategy for tech teams or Claude Code for developers suits you better.

Teaching method

This course is entirely hands-on and code-heavy. You spend two days working in your own IDE with Python or TypeScript and build at least three working agents against the Anthropic API during the course.

Every module follows the same rhythm: a short explanation of the concept, a live demo by the trainer, and then you get to work yourself in the practice repository. You are given code templates that you adapt straight away to your own use case, so that by the end of the course you have code you can deploy in your organisation.

Two labs form the core: on day 1 you build a Q&A bot with document grounding, on day 2 a multi-step agent with tool use and evaluation. With a small group of no more than 8 participants, you have plenty of time for code review and for discussing your specific stack with the trainer.

You take the course in a classroom setting on location or online. For the exercises, you receive API credits for the two course days.

Prior knowledge

You have at least two years' experience with backend development in Python, TypeScript or a comparable language. You are comfortable with REST APIs, JSON and async patterns. Prior knowledge of AI or LLMs is not required. You learn the concepts around Claude, tool use and agents during the course.

Comments

Bring your own laptop with a working Python or TypeScript environment. Arrange an Anthropic API key in advance, so that you can work against real endpoints from the very first moment.

In-company

Would you like to organise this training for a whole group or several groups? This can be cost-effective. For example, we can tailor the content entirely to your organisation, and training a whole group is also cheaper.

Are you looking at large numbers of participants or a complete academy? That involves a great deal. Consider planning and communication, for example, or identifying and agreeing the objectives and content of the training. Or differences in level, feedback mechanisms and safeguarding. Fortunately, Learnit has extensive experience in this field (more than 25 years) and we are happy to help you through this process.

Request a quote

Language

The training is given in Dutch as standard. The trainer has a good command of English. English-language course material can be used. With a minimum enrolment of 3 participants, the training can also be given entirely in English.