Agentplex Weekly - Issue #6
Are LM Agents True AI Agents? melody_agents with crewAI & Suno. Husky an OSS LM Agent. Agents with Gemini Flash. Qwen-agent. Meadow Agents. Agent-universe. Mixture of Agents (MoAs). Agent Hospital.
Are LLM-based agents true AI agents? Since Google Brain published the Attention is all you need paper, a vast amount of research and startups investment has been focused in LLMs and Multimodal LLMs, to the detriment of many important areas in AI/ ML / DL. The Transformers boom has significantly contributed to the rise of LLM-based, or simply LM Agents, while the focus in other interesting types of AI agents like Reinforcement Learning (RL) Agents has somewhat stayed flat. In parallel, there is an increasing level of excitement around combining LM agents with smart-contracts, Web3, and token-based tech. In the mix of all this, some hardcore members of the AI community question whether LM Agents are actual AI agents or not.
“Rather than arguing over what is a true agent, we can acknowledge that there are different degrees to which systems can be agentic.” Andrew Ng
Agentic systems and AI agents. Andrew Ng -a pioneer researcher and investor in Deep Learning- just published a great blogpost on What technology counts as an “agent”? and agentic systems. He argues that there’s a grey zone between what clearly is not an agent (e.g. prompting) and what clearly is an agent ( CoT, ReAct, Auto function calling, etc.) “We should consider agentic systems and the spectrum of agentic technologies,” he concludes.
On agentic workflows. Here’s a live presentation from Andrew 3 days ago, talking On AI Agentic Workflows And Their Potential For Driving AI Progress. And if you want to know more, I suggest you read Andrew’s five-part blogpost on Agentic Design Patterns and Agentic Workflows.
Three New LM Agents
Multi-agent music generation. The speed of evolution in GenAI music is mind-blogging. If you haven’t tried them yet, checkout Suno and Udio AI Music generators: simply amazing! This repo contains a crewAI app for automatically generating songs given a topic and a music genre. The song generation is done using the "unofficial" Suno AI API. The app uses 3 agents: a web researcher, a lyrics creator, and a song generator. Repo> melody_agents: When crewAI meets Suno AI.
Husky-v.1 a new, open-source language agent that solves complex, multi-step, numerical, tabular and knowledge-based reasoning tasks. The agent addresses numerical, tabular and knowledge-based reasoning tasks. Husky-v1 uses a code generator, a query generator and a math reasoner as expert models. Checkout the models, the repo and the paper here> Husky: A Unified, Open-Source Language Agent for Multi-Step Reasoning.
Building an agent with Google Gemini Flash model. Gemini Flash is a multimodal LLM model that is very cheap to use and very fast. This is a nice hands-on tutorial showing how to build a customer support agent with function calling. The tutorial also explains why Gemini Flash is surprisingly good for Agents and Function Calling.
Three Open-source Multi-Agent Frameworks
Qwen-agent. This is a very comprehensive framework for developing LLM applications with the tool usage, planning, and memory capabilities of Qwen2 model. It also comes with example applications such as Browser Assistant, Code Interpreter, and Custom Assistant. Checkout the repo here: Qwen-agent.
Meadow. In the real enterprise world, data workflows are a nightmare! Meadow is an agentic framework for building multi-agent data workflows with LLMs with interactive user feedback. Meadow’s approach is to chain several specialised agents like Text-to-SQL, Planner, Executor, Schema Cleaner, Validator, Router agents to perform an end to end data workflow. Blogpost and repo here: Introducing Meadow: LLM Agents for Data Tasks.
Agent-universe. Open-sourced by Alibaba Research, this is a powerful framework for developing apps powered by LLM-based multi-agents. The framework comes with two pre-installed multi-agent collaboration patterns which have been proven effective in real business scenarios: 1) PEER pattern: Plan, Execute, Express, and Review, to achieve a multi-step breakdown and sequential execution of a complex task, and 2) DOE pattern: A Data-fining agent, Opinion-inject agent, and an Express agent that work together to solve data-intensive tasks. Checkout the repo here: agentUniverse.
Two Interesting Research Papers on Agents
Mixture of Agents. Mixture-of-Experts (MoEs) models are a type of Transformer models that perform faster pre-training and inference. This a great intro to MoEs blogpost by the Hugging Face team: Mixture of Experts Explained. Based on the MoE model approach, researchers at Together.ai just introduced Mixture-of-Agents (MoA). MoA is a layered architecture in which each layer comprises multiple LLM agents. Each agent takes all the outputs from agents in the previous layer as auxiliary information in generating its response. The researchers claim MoA beats GPT-4o in several benchmarks: Paper: Mixture-of-Agents Enhances LLMs Capabilities
Agent Hospital. Inspired by the groundbreaking Smallville paper (25 generative agents living in a town) by DeepMind & Stanford, this is a fascinating paper that introduces a simulation of a hospital called Agent Hospital that simulates the entire process of treating illness. All patients, nurses, and doctors are autonomous agents powered by large language models (LLMs). It also introduces MedAgent-Zero, a doctor agent that learns how to treat illness. The evolved doctor agent achieves a SOTA accuracy of 93.06% on a subset of the MedQA dataset. Paper: Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents.
The Agentplex AI Agents Challenge (link)
Are you building an AI agent? Compete in the AI Agents Global Challenge. It’s easy: Simply submit the details below by clicking here: submissions.
Project overview: Briefly describe your AI agent, its objective, and the problem it aims to solve
Technical details: Provide an overview of the technical aspects, models, agentic patterns, and tech/tools used in your AI agent
Use cases: Describe one or more real-world use cases your AI agent addresses
Link to a demo (if you have one)
GitHub repo link (optional): Provide the link to the public GitHub repository containing the code for your AI agent
Reminder: We plan to extend the submission deadline beyond the original date on 1 September. So you can still keep refining your application up to the final submission deadline to be announced in July.
Next month, we’ll announce important further details, including a better allocation of the $1M prize pool, more benefits for the participants, and some guidelines and ideas.
To receive updates and further announcements, keep reading this newsletter, join our Discord channel, follow us on Agentplex X channel, or simply drop us an email.
Upcoming AI Agents meetup in London, July
Next month, we’re planning to organise a meetup in London. If you are interested in giving a talk or do a demo at a meetup, please contact Carlos here.
If you are interested in a summer internship, please submit an email here to apply.
Thank you for reading Agentplex newsletter. Have a great day.
Carlos, another great, informative issue. Thank you. 🙏