About the Seminar
This workshop is a hands-on introduction to building production-oriented AI agents in Node.js using the Claude Agent SDK. Participants move from a single model call to a fully featured agent with custom tools, file and shell access, permission controls, subagents, skills and session persistence.
By the end of the day, each participant will have a running agent in a Node.js project that performs a real multi-step task autonomously and can be handed to a colleague to run.
Who Is This Seminar For?
- js and full-stack developers
- Backend engineers adding AI features to existing services
- Technical leads prototyping agent-based products
- Developers moving from raw API calls to an agent framework
Prerequisites
Required Knowledge
- Basic JavaScript and TypeScript: async/await, promises, modules (ESM)
- js fundamentals: npm, package.json, environment variables
- Ability to read and write JSON and work with typed function signatures
- Basic command-line proficiency (git, npm, running scripts)
Required Setup
- Laptop with admin rights to install software
- js 24 LTS or later and npm installed
- Git installed and configured
- Cursor or VS Code or an equivalent editor
- An Anthropic API key with available credit
- A terminal the participant is comfortable working in
Key Topics
What an Agent Actually Is
- From single completion to the agent loop: observe, decide, act, repeat
- SDK vs. raw Messages API: what the SDK gives you for free
- Anatomy of the Claude Agent SDK for TypeScript/Node
First Agent in Node.js
- Project scaffolding, installation, and API key handling
- The query interface: prompts, options, and streaming results
- Reading the message stream: assistant turns, tool calls, results
- System prompts and setting the agent's role and boundaries
Tools - Built-In and Custom
- Built-in capabilities: file read/write, shell, search, and web access
- Defining a custom tool: name, description, input schema, handler
- Writing tool descriptions, the model can actually use
- Returning results and errors the agent can recover from
- Connecting external capability via MCP servers
- Lab: build two custom tools and have the agent chain them
Control, Safety, and Permissions
- Permission modes and approving or denying tool calls
- Allow-lists, deny-lists, and scoping filesystem access
- Hooks: intercepting the loop before and after tool execution
- Cost and turn limits; stopping a runaway agent
- Lab: add an approval gate and an audit log to the agent
Scaling the Agent
- Subagents: delegating isolated work and keeping context clean
- Structured output and passing results to downstream code
- Managing context: compaction and what to keep
- Lab: split the agent into a coordinator plus two subagents
Testing and Deployment
- Debugging: reading the stream, logging, and reproducing runs
- Testing strategies for non-deterministic behavior
- Wrapping the agent in an HTTP service
Expected Outcomes
- A complete Node.js agent project with custom tools, permissions, and subagents
- Reusable patterns for tool design and safe agent execution
- A working local setup ready to extend against the participant's own systems
- Reference code and a deployment checklist