Shaman AI

We help you develop complex auto-LLMs

🌳 Collaborate with us 👇

[Shaman AI @shaman_ai]

1/ 🧵 In the fast-paced world of large language models (#LLMs) and multi-agent systems, projects like @CamelAIOrg, #GenerativeSimulacra, #GPTeam, #BabyDeerAGI are crafting their own codebases. They hint at a unifying framework…

Enter #LLAMBDAO 👇

[Cyrus @cyrusnewday]

I'm thrilled to announce the alpha release of my labor of love — Llegos — a domain specific library for exploring the design space multi-agent systems!

One of our early users called it "more of a PyTorch, less of a Keras".

What's the deal with Llegos?

[Shaman AI @shaman_ai]

@cyrusofeden What a great opportunity 🙏

[Cyrus @cyrusnewday]

There are many reasons why you’ll want to join my AI Pioneer course.

Over the span of 4 weeks, I’ll be guiding you to:

1. Learn the skill of observing a modelling agent systems
2. Designing and developing agent systems
3. Contribute to AI research with 2 new systems

How? 🧵

[Shaman AI @shaman_ai]

@TheUnicornist @cyrusofeden Hallucinations cause a key error which is caught by the retry handler

[Shaman AI @shaman_ai]

Diagram of our Agent Architecture

https://t.co/9NtEUFBFig

[Shaman AI @shaman_ai]

@TheUnicornist Wdym chaos? Can you illustrate the example further? Not sure I understand.

Also the repo behind this is in maintenance mode since @cyrusofeden has been working on llegos

[Cyrus @cyrusnewday]

Llegos wants to let you focus on what matters.

Check out how I used it to implement a standard multi-agent contracting protocol.

Part 1: Implementation https://t.co/AIWMfuZOgn

Part 2: Output and Sneak Peak of What’s Next https://t.co/Uy0mV2Ktsq

[Cyrus @cyrusnewday]

LLMs can be hard to chain or compose, especially because the output is unstructured at every step.

Llegos is a Python DSL for building strongly-typed, message-passing concurrent agents with LLMs.

It makes building multi-agent systems fun! Looms below 🌱

[Cyrus @cyrusnewday]

What if agents could communicate with each other through strongly typed messages?

The design space for multi-step agent interactions is huge!

Here’s a peek at what’s coming https://t.co/OdjHAdfn17

[Shaman AI @shaman_ai]

Why aren’t more people live streaming their research?

[Cyrus @cyrusnewday]

Inspired by @realGeorgeHotz, I'm doing a live Zoom right now of the first LLM implementation of contract net, a multi-agent coordination pattern.

Link at the bottom of this thread 🌱

What’s a coordination pattern?

[Cyrus @cyrusnewday]

Inspired by @realGeorgeHotz, I'm doing a live Zoom right now of the first LLM implementation of contract net, a multi-agent coordination pattern.

Link at the bottom of this thread 🌱

What’s a coordination pattern?

[Jerry Liu @jerryjliu0]

Any LLM developer building agents will need to build for a multi-agent future.
🤔 Building a single agent that can do a ton of things is hard
💡 Having agents specialize in different tasks - much easier for each agent to do its job well!
🚀 Build hierarchical agents: outer agent focuses on task orchestration -> specialized mini agents do ground work

In this hierarchical world, the API interface between agents becomes crucially important:
1️⃣ Each specialized agent needs to expose a simple but expressive interface
2️⃣ We need to reduce the cognitive burden on the outer agent - let parameters be optional!

We thought deeply about these concepts in building an integration between @llama_index data agents with @zapier's NLA interface.
✅ Hierarchical agent interface: Zapier can process a natural language string
✅ Make required parameters optional: If outer data agent doesn't know the answer, Zapier API can infer it!

Big shoutout to @zapier team for collaboration on this.

Check out our full blog post here for details and learnings! https://t.co/96GsEGqVwj

You can check out the Zapier integration here: https://t.co/LM3meVdn1A

[Shaman AI @shaman_ai]

@altryne @jerryjliu0 @mlejva @CyrusOfEden has been cooking up a library for multi agent systems, and it’s being used for some exciting new research at @CamelAIOrg

The building blocks are there… would love to see this accelerate the field :)

https://t.co/BFXg7TmXC7

[Shaman AI @shaman_ai]

@srush_nlp @cyrusofeden is in there :)

[Cyrus @cyrusnewday]

A concept inspired by @DrJimFan and @jxnlco —

LLM-generated self-healing Python Functions 

A developer LLM and pytest critic can be used to generate the function, and can further be used to heal the function in production, like when the markup changes in this example:

[Shaman AI @shaman_ai]

@uke__uke 🫶

[Shaman AI @shaman_ai]

1/ 🧵 In the fast-paced world of large language models (#LLMs) and multi-agent systems, projects like @CamelAIOrg, #GenerativeSimulacra, #GPTeam, #BabyDeerAGI are crafting their own codebases. They hint at a unifying framework…

Enter #LLAMBDAO 👇

[Shaman AI @shaman_ai]

What would be more helpful? / Exciting? / Cool?

[Shaman AI @shaman_ai]

.@CyrusOfEden just pushed some updates to https://t.co/z6s5yUyUSt:

1. Added docstrings
2. Moved some examples into tests and added documentation around texts for quick starters
3. Renamed SwarmNode to ApplicatorNode for clarity
4. Fixed Actor class

[Shaman AI @shaman_ai]

2/ 🧠 LLAMBDAO reimagines multi-agent systems as networks of Nodes, communicating via Messages. It's a language for exploring agent architectures.

Fun fact: LLAMBDAO is an abbreviation of “Large Language Agents Modulating Behaviors in Decentralized Autonomous Organizations”

[Shaman AI @shaman_ai]

And a hint of the helpers for working with OpenAI — particularly, converting nodes to schemas that can be called with the function call API.

https://t.co/TAd0bb47ph

[Shaman AI @shaman_ai]

8/ 🌱 We present LLAMBDAO as a work-in-progress, rather than a final product. We invite you to join us in this exciting exploration, to push boundaries, question assumptions, and shape the future of autonomous agent architectures.

https://t.co/C5it6ENZY9

[Shaman AI @shaman_ai]

7/ 🌍 We're at the start of a fascinating journey with autonomous agents. LLAMBDAO isn't the destination—it's a journey, a language, and a collaborative venture.