Claude Agent SDK [Full Workshop]
summary
Workshop shows an implemented Agent Loop (Context → Thought → Action → Observation) inside a built Harness that wires tools, prompts and a file system for persistent context; it treats the Bash tool as the primary generic actuator, demonstrates sandboxing with ReadOnly vs ReadWrite permissions, and uses Hooks for deterministic checks. Live code generates a TypeScript PokéAPI SDK, runs agents via Bun, exposes a real-time monitor of thoughts, and recommends Claude MD files and subdirectory starts for 50M+ line repos.
AI-written summary · about
chapters (16)
- 0:00Introduction: Agenda and the "Agent" definition
- 5:15The "Harness" concept: Tools, Prompts, and Skills
- 10:10Live Coding Setup: Initializing the Agent class and environment
- 15:45implementing the "Think" step: Getting the model to reason before acting
- 25:20The Agent Loop: connecting `act`, `observe`, and `loop`
- 33:10Tool Execution: Handling XML parsing and tool inputs
- 42:00The "Bash" Tool: Giving the agent command line access
- 49:30Safety & Permissions: "ReadOnly" vs "ReadWrite" file access
- 58:15Context Engineering: Using `ls` and `cat` to build dynamic context
- 1:05:00The "Monitor": Viewing the agent's thought process in real-time
- 1:12:45Handling "Stuck" States: Feedback loops and error correction
- 1:21:20Multi-turn Complex Tasks: Building a "Research Agent" demo
- 1:35:10Refactoring patterns: "Hooks" and deterministic overrides
- 1:48:39Q&A: Reproducibility, helper scripts, and non-determinism
- 1:50:31Q&A: Strategies for massive codebases (50M+ lines)
- 1:52:00Closing remarks and future SDK roadmap
related talks
44mHow Does a Coding Agent Work?
01h 21mFull Walkthrough: Writing & Using Skills
01h 06mHow I Tamed Claude
18mClaude Code & the evolution of agentic coding
11mHow to build a Terminal Agent like Claude Code or Codex
39mHow to build a Terminal Agent like Claude Code or Codex
46mVibecoding your Own Multi-Agent Workstation
16mBeyond the IDE: The Future of Developing Software With Agents