The Claude Code SDK is now the Claude Agent SDK Why? Because we realized the Claude Code agent harness is useful for much more than coding. In fact, we're moving to using it to power most of our own agent loops at Anthropic.

The key insight behind the Claude Agent SDK is that by giving Claude access to a computer, it can do work like humans do. We think this should help create a new wave of agents in finance, customer support, research, etc. that are as effective as Claude Code is at coding.

Building an agent loop with the SDK is simple- your agent needs to be able to gather context, take action and verify its work. But building simple things is hard, you'll have to carefully pay attention to how your agent operates and be creative with what tools it has.

An often useful tool when building your agent loop is to use the file system often. I wrote more on how you can here:

If you want to see an example, I’m building an email agent with the Claude Code SDK in public. It’s still a work in progress but you can learn more here:

Read the full article with more examples here:

memory function in agent sdk wen? looks like claude has generalized on using something called mcp__memory__ in a similar way. but to get the full natively trained memory tool call behaviour the tool has to be called "memory", right?

The memory tool is file system based, you can implement your own version fairly quickly with the SDK. But we are thinking about how we can make this more native.

wdym? api docs say you need to implement the memory tool function client side following the predefined schema. 4.5 was RL trained on using this 'memory' tool natively, no?

wen shitty node ui

would love a better explanation and diagram on the claude sdk deployment process and dependencies. the docs are missing clarity around claude cli dependency and access to file system on prod env resulting in multiple error iterations on first deploys.

Could you tell us how to use claude agent sdk with max subscription? I dont want the pre loaded system prompts from claude code. Any docs about that?

plans on releasing a UI like the agent builder? lower the barrier for non-technical people that dont want to use a CLI

Just wrap a couple of commands around Claude SDK, you can make all kinds of SaaS out there, small to large. Seems to me all of them. Only concern is token cost.

Beautiful but what’s the point on build such beautiful tools if weekly limits won’t let you use them? Lots of users in higher plans are complaining about and not a word from Anthropic.

How do you manage the runtime environment for the agent? Give each agent it's own filesystem with bubblewrap or something similar?

Yeah that's right, or use a sandbox provider like Daytona, E2B, etc.

Is deep research powered by it? My usage of deep research recently shifted from OpenAI to Claude

Exciting evolution for Claude! Expanding from code to broader agent capabilities like finance and research opens up so many real-world applications looking forward to seeing how devs build with this.

Excited to see how expanding the SDK's capabilities will enhance AI agent development and applications. Looking forward to the innovations this will bring.

How do I have it connect to a remote dev container to work in it rather than local file system?

The most important part of this announcement isn't the name change. It's the signal that Anthropic is running its own production agent loops on the same harness you're giving to the public. That's the ultimate endorsement.

model-environment co-design 👀

Love all the work your team is doing. It’s always fun to dive into new updates and docs and start applying :)

Would be interesting if you guys put together more agents that can be created in Claude code that aren’t software dev use cases. I get that that’s a large TAM but most people aren’t devs. Democratizing Claude code for non-coding use cases would be great

A similar thing google did - gemini sdk , became a google agent toolkit

The rename signals a strategic shift that most people are underestimating. "Code SDK" implied developer tooling. "Agent SDK" signals infrastructure for autonomous systems — orchestration, permissions, human confirmation gates, production readiness baked in. The fact that Anthropic is now using it to power their own internal agent loops means this isn't just an external product. It's their own production architecture made available to builders. That's a very different level of confidence in the stack.

Code was a fun stop. Every task-specific SDK is just a temporary alias on the road to generalized Agent architecture. Soon, Agent will be dropped too. Next up: The Just Do Everything SDK.

This has me hyped up since release, but one thing I don’t see much talk about is how do you actually deploy an agent to operate online? I love how CC moves through local files for context, but how does that translate to a server/db? I’m not technical, so excuse my vocabulary lol

For devs, isn't it more useful to just use headless claude code? Web search, file system writes, etc are already included, and you can reproduce functionality that the agent sdk offers very quickly. This can all fall under max subs instead of paying huge api costs

authentication via claude ai subscription not via api key when?

Serious question for you Thariq. Since Claude Agent SDK and your proposals involve the file system and OS level execution, what runtime env are you running these agents loops in production at Anthropic?

Claude Agent SDK is awesome!! But I’m having issues with the tool permissioning and seems to be a bug since the examples code doesn’t work

So... You can basically "agentify" anything with this? Beyond just building software.

I need to spend a couple days with the SDK! BTW, is there a fix coming for the sessions terminating in the CC plugin. I keep having to login with CC

Claud max plan usage has been silently reduced 20x why? Because most unethical company realised that they can rip off customer by scheming and not communicating properly.

Would be much more useful if it would be open source

When will it be model agnostic :)

When better documentation?!?!?

Excited for the broader applications of the Claude Agent SDK. Curious about the challenges you faced while integrating its capabilities outside coding. Any lessons learned so far?

An article about Claude code agent use cases would be extremely,extremely interesting …

how to create this video ？