Agentplex Weekly - Issue #10
New Agentic Interfaces for Billions of AI Agents. Large Action Models. Agentic Long Term Memory. Agentic RAG. ReAgent. Einstein Autonomous Agent. Multi-Agents Play Online Games. DiagnosisGPT.
Agentic Interfaces for billions of AI Agents working and consuming. The common view across many investors, researchers and startups working in AI Agents is that soon there will be billions of AI Agents in the wild. Billions of AI Agents that will operate both as “autonomous workers” executing all sorts of jobs, and as “autonomous consumers” buying all kinds of services. But this will require radically new ways on how humans interface with AI agents, and how AI agents interface with other AI agents. And also understand how much autonomy we should allow in those agents.
To support the rise of billions of AI Agents, first we will need a new, solid, robust, reliable AI Agents infrastructure tailored to the very specifics of how AI agents will interact with agents and humans in the future; just scaling cloud GPU compute won’t be enough. As Emad the founder of Stability AI writes, all these billions of AI Agents will need some basics, like for example new payments infra to get paid for work they will execute, and to pay for services they will consume.
All websites dynamically generated by AI Agents. Jeremiah Wang - the startup investor and one of the judges of The AI Agents Global Challenge - recently wrote an interesting article on what will happen when AI agents dominate the Internet and outnumber humans users. His main point: All existing websites will become obsolete, and AI Agents will dynamically generate the website content when a human visits a website on demand. This will require new types of human-agents UIs. Blogpost: My vision: The Rise of AI Agents.
New AI agent-human interfaces based on intentions. Jeremiah’s post brings me to the concept of Large Action Models (LAMs). A LAM is a type of AI agent that is powered by a foundation model pre-trained on billions of human-machine UI actions. LAMs focus is on understanding tasks based on standard interfaces like web browsers using Vision-Language capabilities. The main goal of LAMs is to predict user intentions and automatically execute a task accordingly. LaVague is a very popular LAM, and Rabbit OS is one of the first “commercially available” LAM. This is a great blogpost on: what LAMs are, how they work and the difference between LAMs and LLM-based Agents.
Agent-Computer Interfaces and Billions of AI Agents as consumers. In this post, the author argues that having billions of AI Agents as consumers will be a problem initially, as this will drive the Internet supply-demand forces towards negative-sum outcomes. Eventually, “agentic-market dynamics” and the economic benefits of Web 4.0 will kick in, and positive outcomes will emerge. The level of AI agents autonomy will be key too. And to enable all this, there will be a need to develop new agent-computer interfaces (ACI) Blogpost: The Path To Autonomous AI Agents Through Agent-Computer Interfaces (ACI)—Onward To Web 4.0.
Hands-on, tutorials and practical guides
Building a multi-agent banking concierge. This blogpost shows how to use LlamaIndex to build a multi-agent banking system with agents responsible for each top-level task, plus a "concierge" agent that can direct the user to the correct agent.
How to enable AI Agents with long-term memory. This tutorial shows how to implement multi-level and long-term memory in AI agents using mem0, running locally and with cross-platform consistency.
How to build a RAG Agent. This video tutorial shows how to build a RAG assistant on you local machine using Nomics' GPT4All, Llama 3 models with Google Vertex AI Agent Builder. Also checkout this vid on Agentic RAG by Jeremy CEO at LLamaIndex.
New AI Agents, tools, platforms, and frameworks
Agentic Node Graph Generator. This is an agentic workflow system that generates node graphs for workflow automation apps like Magpie, Zapier and ComfyUI. This system is generally applicable to node-graphs or DAG's.
Reagent, graph based framework for full-stack AI agents. This is an open-source Javascript framework to build AI workflows. It enables you to build multi-step workflows by combining nodes into an agentic graph.
Salesforce Einstein Service Agent. Salesforce claims that this is the first fully autonomous customer service AI agent that understands context and nuance, and is intelligent and dynamic enough to autonomously determine the next actions to take.
AI Agents Startups
Henry.ai - AI assistants for real-estate
Bryter - AI Agents for legal & compliance
Marr Labs - AI voice agents indistinguishable from humans
Thoughtful.ai - fully human-capable AI agents for healthcare
Research Papers
A multi-agent foundation model for playing online-games. This project -checkout the paper, code & demo- introduces Cradle, a General Computer Control (GCC) agents framework. Cradle to ace any computer task by enabling strong reasoning abilities, self-improvement, and skill curation.
Agents for complex, real-world reasoning. This paper introduces Sybyl, a simple yet powerful LLM-based agent framework designed to tackle complex reasoning tasks by efficiently leveraging a minimal set of tool.
DiagnosisGPT: An interpretable, AI Agent for medical diagnosis. This paper introduces Chain-of-Diagnosis (CoD) and DiagnosisGPT, an agent-based medical diagnostics that mirrors a physician's chain of thought process. DiagnosisGPT is capable of diagnosing 9,604 diseases.
Are you building an AI Agent?
Compete in the AI Agents Global Challenge funded with a $1 Million pool prize. It’s easy to apply, with fairly broad scope, and there is no registration fee. Click here to learn about the challenge and how to apply.
Join our Discord channel to meet other people building AI agents, discuss ideas and collaborate in projects.
Thank you for reading Agentplex newsletter. Have a great day.