Agent SDK shipping

The harness Claude Code is built on.

If you've used Claude Code and wondered "how is it doing all this looping," the answer is: the Agent SDK. You can build your own version on the same infrastructure — a goal, some tools, a budget, and Claude runs until done.

What you get

• An agent loop that runs until it decides it's finished (or hits a step budget).
• Built-in tools — file read/write, bash, web fetch, glob, grep — opt in or out.
• Sub-agents, so a planner can spawn workers.
• Persistent memory and conversation state.
• Context-window management — the SDK summarizes earlier turns when the window fills.

Different jobs.

From zero to a running loop.

1. Install: pip install claude-agent-sdk or npm install @anthropic-ai/claude-agent-sdk.
2. Write a goal as a plain-English instruction: "Read every .md file in ./docs, find broken links, write a report to ./report.md."
3. Pick the tools the agent is allowed to use (file tools yes, bash maybe, web fetch maybe).
4. Run it. The SDK loops, calls tools, reports back when done.

from claude_agent_sdk import Agent
agent = Agent(
    model="claude-sonnet-4-6",
    tools=["read_file","write_file","glob","grep"],
    goal="Find broken links in ./docs and write a report to ./report.md",
    max_steps=40,
)
agent.run()

Three shapes you'll reuse.

Don't ship without these.

• Step budget. Always cap max_steps. A runaway loop wastes money fast.
• Permissioned tools. Whitelist explicit folders for file tools. Never give bash on a daily-driver machine without thinking.
• Token telemetry. Log response.usage per step. Dashboards in week one.
• Resume on failure. Persist agent state so a crash can pick back up.
• Human approval gates for irreversible actions — push, deploy, send-mail, destructive SQL.

Related products.

Every Anthropic surface has its own page on this guide.