Agent SDK Deep Dive

Steps

1. Agent SDK — Overview anthropic-platform intermediate

   Introduction to Anthropic's Agent SDK for building custom AI agents with tool use, MCP, and sub-agents.

   Start here to understand what the Agent SDK provides beyond the raw Messages API. It handles the agentic loop, tool execution, and multi-turn orchestration that you would otherwise build yourself.

2. Quickstart anthropic-platform beginner

   Get a working agent running in minutes. Focus on the AgentConfig pattern since it is the central abstraction — everything else (tools, subagents, hooks) plugs into this configuration object.

3. Typescript anthropic-platform intermediate

   The TypeScript SDK uses generics extensively for type-safe tool definitions. If you are a TypeScript shop, start here instead of the Python guide — the APIs are equivalent but the type system catches configuration errors at compile time.

4. Python anthropic-platform intermediate

   The Python SDK is async-first using asyncio, so make sure your runtime supports it. Python typically receives new Agent SDK features first, making it the safer choice for accessing the latest capabilities.

5. Custom Tools anthropic-platform intermediate

   Tools are what make agents useful beyond conversation. The quality of your JSON Schema definitions directly determines how reliably the agent invokes your tools — invest time in clear parameter names and descriptions.

6. Mcp anthropic-platform intermediate

   MCP integration lets your agent use any compatible tool server without writing custom tool code. This is the extensibility story — read this after custom tools to understand when to build tools inline versus connecting an MCP server.

7. Subagents anthropic-platform intermediate

   Subagents enable delegation patterns where a coordinator agent farms out tasks to specialized workers. Compare this with OpenAI's handoff pattern — Anthropic uses delegation with context passing while OpenAI uses explicit transfer of control.

8. Streaming Vs Single Mode anthropic-platform intermediate

   Single-mode is simpler to implement and debug, while streaming gives users real-time feedback. Start with single-mode for development, then switch to streaming for production UIs — the migration path is straightforward.

9. Streaming Output anthropic-platform intermediate

   This covers the event types you will receive during streaming execution. Understanding the event lifecycle (turn_start, tool_use, tool_result, turn_end) is essential for building responsive agent UIs that show progress.

10. Hooks anthropic-platform intermediate

    Hooks let you intercept agent behavior at key lifecycle points — before/after tool calls, on errors, and at turn boundaries. They are the right place for logging, guardrails, and approval workflows without modifying agent logic.

11. Permissions anthropic-platform intermediate

    Permissions control what an agent can do at the tool level. This is critical for production — develop with permissive settings but deploy with least-privilege, or a prompt injection could escalate through your tool surface area.

12. Hosting anthropic-platform intermediate

    Agent runs can take seconds to minutes, which is fundamentally different from typical API request durations. Focus on the timeout and concurrency considerations — these affect your infrastructure choices more than anything else.

13. Secure Deployment anthropic-platform intermediate

    Agent security goes beyond API security because agents autonomously execute tools. Focus on the prompt injection mitigations and input validation guidance — these are the primary attack vectors for deployed agents.