Anthropic · Technologies
Claude Agent SDK
The engine behind Claude Code, in your hands.
The Claude Agent SDK is Anthropic’s open source library for building production agents in Python and TypeScript. It exposes the same agent loop, tools and context management that power Claude Code, so your agents can read files, run commands, search the web and drive MCP connections out of the box. We build production agent systems on it, including our own.
AGENT FLEET · EXAMPLE SYSTEM
AGENT
ROLE
HANDOFF
STATUS
Research-Agent
Gather
To writer
LIVE
Drafting-Agent
Produce
To review
LIVE
Review-Agent
Check
To human
LIVE
Hooks
Guardrails
Enforced
REVIEW
rogue-experiment
Unknown owner
None
RISK
Sample system
live · review · risk
In plain terms
One agent is a demo. A governed fleet is a system.
The gap between an agent that impresses and one that ships is engineering. Here is what changes.
Without it
- The agent loop invented from scratch per project
- Tool execution, permissions and context managed by glue code
- No control over what an agent may and may not do
- Experiments multiplying with no owners
With it
- A proven harness doing the loop, tools and context
- Subagents splitting big jobs with their own scoped tools
- Hooks enforcing your rules at every step
- A registered fleet with owners and guardrails
What CG TECH can do with the Claude Agent SDK
The work, broken into the parts that matter.
The harness Claude Code runs on
The SDK is the same engine that powers Claude Code, hardened by daily use at enormous scale, exposed as a library your team calls from plain Python or TypeScript.
a harness proven in production
Control built in, not bolted on
Subagents delegate work to child agents with their own context and tool set, and lifecycle hooks let you validate, log or block any action before it happens.
agents with rules they cannot skip
Your systems, one protocol away
The SDK speaks Model Context Protocol natively, so the connections you build to your systems work here, in Claude and across the growing MCP ecosystem.
connect once, reuse everywhere
The right home per workload
Run agents in your own environment with the SDK, through your cloud via Bedrock, Google Cloud or Microsoft Foundry, or on Anthropic’s hosted Managed Agents service. We recommend per workload.
agents where they belong
How an engagement runs
From a demo agent to a governed fleet, step by step.
01
Discover
We map the workflow, the systems involved and the risk profile.
02
Design
Agent roles, tools, hooks and guardrails designed before code.
03
Build
The system shipped with logging and evaluation from day one.
04
Handover
Code, runbooks and an operating rhythm your engineers own.
Questions we hear a lot
Common questions about the Claude Agent SDK
What is the Claude Agent SDK?
The Claude Agent SDK is Anthropic’s open source library for building production agents in Python and TypeScript. It exposes the same agent loop, tools and context handling that power Claude Code, so an agent you build can read files, run commands, search the web and drive MCP connections without you writing that plumbing yourself.
Claude Agent SDK or the OpenAI Agents SDK?
They solve the same problem on different models, and we build with both. The Claude Agent SDK ships with deeper built in tools and MCP support; we recommend per workload and estate, honestly.
Do we need our own infrastructure to run agents?
No. Agents can run in your environment, through your existing cloud, or on Anthropic’s hosted Managed Agents service where Anthropic operates the harness. We match the deployment to your governance needs.
How do we keep autonomous agents safe?
Scoped tools per agent, hooks that block dangerous actions, human approval on consequential steps and full logging. Autonomy is earned in stages, not granted on day one.
Is it locked to Anthropic?
The SDK is open source and built for Claude models, which you can reach directly or through AWS, Google Cloud and Microsoft Foundry. The MCP connections you build alongside it are an open standard and portable anywhere.
What is an agent harness?
It is the machinery around the model: the loop that decides what to do next, the tools it can call, and the management of what stays in context. Most of the difficulty in building a reliable agent sits there rather than in the model, which is why starting from a proven harness saves months.
What can an agent built on it actually do?
Read and write files, run commands, search, and reach your systems through MCP. In practice that means work like processing a queue of documents, reconciling data between two systems, or running a multi-step research and drafting job end to end.
Who maintains it after you build it?
Your team, and we build it to be handed over. Agents need an owner in the same way applications do, because the systems underneath them change. We document what it does, how it fails and who to call.
Ready when you are
Building a real agent system? Let us talk.
A discovery session maps your workflow, your risks and your quick wins. You keep the plan either way.
What to expect
- A consultant replies within 4 business hours
- Session booked to understand your requirements
- We will provide you with a fixed price quote