Claude Agent SDK [Full Workshop]
Bash is the most powerful agent tool. The Claude Agent SDK packages Claude Code's battle-tested patterns—tools, file system, skills, sandboxing—for building coding and non-coding agents alike.
Overview
Anthropic built the Claude Agent SDK by extracting patterns from Claude Code. Engineers kept rebuilding the same agent harness components: tool loops, file systems, prompts, skills, sub-agents, memory. The SDK packages these lessons into an opinionated framework for building production agents.
Key Arguments
Bash Is All You Need
The bash tool made Claude Code exceptional. It functions as the first "code mode"—enabling programmatic tool composition before formal code mode existed.
What bash enables:
- Dynamic scripts: Generate and execute code on-demand rather than hardcoding tools
- Composability: Pipe outputs, store results to files, chain operations
- Existing software: Use ffmpeg, jq, grep, LibreOffice without building integrations
- Self-checking: Save intermediate results to verify work
Without bash, an email agent searching for "how much did I spend on ridesharing?" must process 100 emails in context. With bash, it can grep for prices, store them to a file, and sum them—checking its work at each step.
Code Generation for Non-Coding Tasks
Claude Code succeeds at non-coding work because it generates code to solve problems.
When asked about weather and outfit recommendations, Claude Code might:
- Write a script to fetch a weather API
- Get location dynamically from IP address
- Call a sub-agent for wardrobe recommendations
- Compose everything into a reusable tool
This pattern—composing APIs via generated code—applies to email processing, data analysis, and document generation. The SDK's Office agent and dashboard builder examples use this approach.
The Three Agent Actions
Every agent operation falls into tools, bash, or codegen:
| Method | Pros | Cons |
|---|---|---|
| Tools | Structured, reliable, minimal retries | High context usage, not composable |
| Bash | Composable, low context, uses existing software | Discovery latency, slightly lower call rates |
| Codegen | Highly composable, dynamic, API composition | Longest execution, needs linking/compilation |
Use tools for atomic, irreversible actions (write file, send email). Use bash for composable workflows (search, lint, memory). Use codegen for dynamic logic and API composition.
The Agent Loop Pattern
Three phases define effective agent design:
- Gather context - Find relevant information (grep files, search emails)
- Take action - Execute work (tools, bash, codegen)
- Verify work - Confirm results (lint, compile, execute)
The verification step determines agent viability. Coding agents verify via linting and compilation. Research agents verify via source citation. Agents without strong verification remain unreliable.
Swiss Cheese Security
Agent security uses layered defenses:
- Model alignment - Training to avoid harmful actions
- Harness controls - Permission prompts, bash parsing
- Sandboxing - Network isolation, file system boundaries
The "lethal trifecta" attackers need: code execution, file system access, and data exfiltration. Network sandboxing blocks the third requirement.
Notable Quotes
"Bash is what makes Claude Code so good. If you were designing an agent harness, you'd have a search tool and a lint tool and an execute tool. Every time you thought of a new use case, you'd need another tool. Instead, Claude just uses grep." — Thariq Shihipar
"The number one meta-learning for designing an agent loop: read the transcripts over and over again. Every time you see the agent running, read it and figure out—what is it doing? Why? Can I help it somehow?"
Practical Takeaways
- Design agents around verification: if you can't verify the output, the agent will struggle
- Use tools for actions requiring user approval; use bash for everything composable
- Skills are folders the agent can read—just files with expertise baked in
- Think context engineering, not just prompts: tools, files, and scripts shape behavior
- For workflows, still use the SDK—even structured tasks benefit from bash composability
Connections
- 12-factor-agents - Shares the philosophy of small, focused agents with deterministic control flow and strategic LLM decision points
- boris-cherny-on-what-grew-his-career-and-building-at-anthropic - Another Anthropic engineer's perspective on Claude Code's architecture and impact
- understanding-claude-code-full-stack-mcp-skills-subagents-hooks - Explains the extensibility layers that the Agent SDK builds upon