Overview
Thariq Shihipar from Anthropic presents a comprehensive workshop on the Claude Agent SDK, explaining how it builds on Claude Code to create powerful agents. The key insight is that bash is the most powerful agent tool - it enables agents to compose functionality, store results dynamically, and work with existing software rather than being limited to predefined tools.
Key Takeaways
- Start prototyping with Claude Code directly - test your agent ideas by giving Claude Code your APIs and custom prompts before building production systems
- Use the file system for context engineering - store memories, scripts, and data in files that agents can read and modify, rather than cramming everything into prompts
- Design agents around the gather context → take action → verify work loop - focus on making each step as robust as possible with deterministic verification rules
- Bash enables true composability - agents can pipe outputs, store intermediate results, and combine multiple tools dynamically rather than being limited to predefined tool combinations
- Think about reversibility when choosing agent domains - agents work best on problems where mistakes can be undone (like code with git) versus irreversible actions
Topics Covered
- 0:00 - Introduction and Workshop Overview: Introduction to Claude Agent SDK workshop agenda covering what it is, why use it, and live coding session
- 2:00 - Evolution of AI Features: How AI has evolved from single LLM features to workflows to autonomous agents like Claude Code
- 4:00 - What is Claude Agent SDK: SDK components including models, tools, prompts, file system, skills, and how it packages common agent building blocks
- 8:00 - The Anthropic Way to Build Agents: Opinionated approach using Unix primitives, bash tools, file systems, and code generation for non-coding tasks
- 15:30 - Bash is All You Need: Deep dive on why bash is the most powerful agent tool - enables composability, memory, dynamic scripts
- 21:30 - Agent Loop Design: Three-part agent loop: gather context, take action, verify work, with emphasis on reading transcripts to improve
- 25:30 - Tools vs Bash vs Code Generation: When to use each approach - tools for atomic actions, bash for composable operations, codegen for dynamic logic
- 30:30 - Skills and Progressive Context Disclosure: Q&A on skills as folders of expertise, how they work with file system, and best practices
- 50:30 - Designing a Spreadsheet Agent: Interactive exercise designing agent search interfaces, API transformations, and verification strategies
- 1:23:30 - Live Coding: Pokemon Agent: Prototyping session building a Pokemon agent with PokeAPI, comparing tools-only vs Claude Code approaches
- 1:39:00 - Advanced Agent Examples and Q&A: Competitive Pokemon team building, deployment strategies, monetization, and handling large codebases