We have just released the Letta Agent SDK, and I hope you will try it out!
Letta agents are stateful, meaning that they have the opportunity to learn with every invocation -- and now, you can provide those opportunities with a sane + clear SDK.
The Letta Agent SDK is designed to make it easier for developers/agents to integrate any Letta agent fully with your application (cloud or local).
Some things you can do:
Custom interfaces (we forked Signal Desktop and made it a Letta interface, see also our mobile app demo)
Dynamic workflows (check out the examples)
White-labeled applications on top of Letta
You can very easily stand up an agent in any aspect of your business or software or workflow, and allow that agent to flexibly adapt to the situation you've deployed it in.
For example, you could create a small piece of code that fetches papers from the arXiv and pass each paper to an agent. Because the agent is a Letta agent, it passively learns through the act of doing.
This extends to all aspects of agents. You can deploy them in Linear, and the agents will just passively start understanding Linear better because of how memory works in Letta.
As a side note -- dreaming is completely free if you use Letta Cloud. Pro-tip.
Agents are also capable of extending themselves because agents can write Agent SDK code to build dynamic workflows -- something like appendages or arms to hook themselves into the broader world around them.
I have used the Agent SDK to build automated triage workflows that forks our primary software engineering agent using a cheaper model. These forks have all the same memory and ability to learn, but can be run at a much larger scale and lower cost.
The Letta Agent SDK will feel familiar to people who have used the Claude Agent SDK.
We adapted magnificent work from the Anthropic team on the Claude Agent SDK, but we've made it stateful, model-agnostic, and work with cloud or local agents. It fully supports using any computer or app server you have available.
Resources
I've been experimenting with building on the Agent SDK for a few weeks now, and here are some of my favorite projects:
Hypervigilant, a tool for monitoring changes in files and sending them to agents. It's very cool -- worth a tinker. Check out this demo.
thought stream, my personalized event stream built on Jazz.
How to fan an agent out to learn from documentation.
General tips for building with the Agent SDK.
The history of programming Letta agents
I was originally hired as a developer relations person at Letta because we were targeting developers. For the past six to nine months, we've had a big technical transition away from the legacy server-side model, which you can read about in our blog post. We've finally stabilized a lot of the way that Letta works under the hood so that we can return to programming Letta agents.
I hope that everyone will take this as an opportunity to look at Letta as a first-class place for developers to build applications where you care about state, learning, and customization at the edge. Build agents that learn where knowledge and learning matter.
I genuinely believe that our SDK is one of the most beautiful ways to program agents, and I also believe that Letta's approach to agents is worth your time and investment.
Pride
I am extremely proud of the team and the work we've done, and I'm happy that we provide a first-class agent programming toolkit so that your agents can now be everywhere that code is.
You can check out the Agent SDK on GitHub, or see our magnificent documentation. You can also chat about the Agent SDK on our Discord!
I would love to see what you build!
-- Cameron