The Claude API has grown from a single chat endpoint into a platform for building agents: client tools you execute, server tools Anthropic executes (web search, web fetch, code execution), tool search for large toolsets, prompt caching and batches for cost, context editing, compaction and a memory tool for long-running work, a connector for remote MCP servers, Agent Skills, and the Claude Agent SDK that packages the same agent loop used by Claude Code.
This course works through each layer in the order you need it in a real build. You start with the Messages API request and response shape, streaming and token accounting; move to the tool-use loop and its failure modes; then cut cost and latency with cache breakpoints, the Message Batches API and file uploads. The second half covers long-running agents: managing the context window, connecting MCP servers, packaging procedures as Skills, building with the Agent SDK's subagents, hooks and permissions, and deciding when Anthropic's hosted Claude Managed Agents harness (in beta as of October 2026) is the better fit. It closes with what changes when you call Claude through Amazon Bedrock, Claude Platform on AWS, Google Cloud or Microsoft Foundry instead of the direct API, and how to evaluate an agent before launch. Product facts are stated as of October 2026; beta features are labeled as such and checked against platform.claude.com documentation.