Claude Agent SDK Development: What It Is & How We Use It
The Claude Agent SDK is the harness behind Claude Code, packaged as a library you can run on your own infrastructure. If you want a coding-style agent that can read files, run commands and search the web without writing the loop yourself, this is the shortest path. Here is what it is and how we use it for clients.
- Published
- Updated
What the Agent SDK gives you
The SDK ships as a Python and a TypeScript package. You give it a prompt and options; it drives the full agent loop with a set of built-in tools — file read, write and edit, shell, glob and grep, web search and fetch — plus MCP connections, subagents, hooks and a permission system. It is the same machinery that runs Claude Code in a terminal, minus the terminal.
| Claude Agent SDK | API + tool runner | Managed agents | |
|---|---|---|---|
| Who writes the loop | SDK | SDK helper (tools you define) | Anthropic, server-side |
| Built-in tools | Files, shell, search, web | None — yours only | Sandbox: bash, files, code exec |
| Where it runs | Your infra | Your infra | Anthropic-hosted container |
| Subagents | Built in | You build them | Multi-agent sessions |
| Best for | Filesystem / coding / research agents | Custom business tools | Hosted, scheduled, long-running agents |
When we reach for it
- Internal engineering agents: codebase Q&A, migration bots, test writers, PR reviewers wired to your repos.
- Research and document agents that need to read a folder of files, search the web and write a report.
- Ops automation where the "tools" are really shell commands and files — log triage, config generation, data cleanup.
- Anything where you would otherwise script Claude Code in CI. The SDK is the supported way to do that.
When the tools are business APIs rather than files — CRM, payments, your product’s backend — we usually go to the plain API with the tool runner instead. It is thinner and you do not carry tools you will never call.
How we ship on it
- Week 1: define the job, the allowed tools and the permission policy. Most agents need far fewer tools than teams expect.
- Week 2: hooks for logging and approval, MCP servers for the systems the agent must touch, first eval set from real tasks.
- Weeks 3–4: subagents where the work fans out, cost tuning with effort levels, hardening against the failure cases the evals expose.
- Handover: your repo, your infra, a runbook and a training session for the engineers who will own it.
We also teach it
Our one-day Agent SDK workshop takes an engineering team from zero to a working agent on their own codebase: loop, tools, hooks, permissions, evals. Teams leave with a starter repo and the judgement to know when the SDK is the wrong tool. Ask for it through the form below.
Frequently asked questions
Is the Claude Agent SDK the same as the Anthropic SDK?
No. The Anthropic SDK is the API client (messages, tools, batches, files). The Claude Agent SDK is a separate package that bundles the Claude Code harness and built-in tools. Both run on your infrastructure.
Can the SDK agent use my internal tools?
Yes — through MCP servers or custom tools registered alongside the built-ins. That is how we connect it to ticketing systems, databases and internal APIs.
Which model should it run on?
For agentic work we default to the current Opus-tier model with adaptive thinking and a high effort setting, and drop cheaper models in for subagents that only read and summarise.
