The Claude Agent SDK: build your own agents

Why an SDK?

Claude Code is an agent for coding. But businesses want agents for support tickets, document processing, research, operations. The Claude Agent SDK exposes Claude Code's engine as a library, so you can build those agents without reinventing the hard parts.

• Python: pip install claude-agent-sdk
• TypeScript: npm install @anthropic-ai/claude-agent-sdk

A tiny agent

import asyncio
from claude_agent_sdk import query, ClaudeAgentOptions

async def main():
    options = ClaudeAgentOptions(
        system_prompt="You are a helpful operations analyst.",
        allowed_tools=["Read", "Grep", "Glob"],
    )
    async for message in query(prompt="Summarise the CSVs in ./reports", options=options):
        print(message)

asyncio.run(main())

Notice: you set the role, restrict tools to read-only, give a task and stream messages back. The SDK runs the loop of think → call tool → observe → repeat for you.

How the agentic loop works

Under the hood (and if you call the Messages API yourself), each response has a stop_reason:

• tool_use → run the requested tool(s), append the results to the conversation, call Claude again.
• end_turn → Claude is finished; return the answer.

Production essentials

• Tool descriptions are how the model picks tools, so make them specific, with inputs, examples and boundaries.
• Few, focused tools per agent: around 4–5 scoped to the role beats 20 generic ones.
• Hooks for business rules that must never be broken (refund limits, approvals).
• Structured errors: tell the agent whether a failure is retryable, and distinguish "couldn't reach the system" from "searched and found nothing".
• Human handoff: when escalating, pass a self-contained summary (customer, issue, what was tried, recommended action).

Check yourself

Three quick questions to check it stuck.

Ask the tutor

Frequently asked