Agentplex Weekly - Issue #11
Agentic Workflows. New Llama 3.1 450B’s Agentic Capabilities. Modern Agentic RAG Pipelines. Improving Agentic Planning. Building a SQL Agent. Agent-E for Web Automation. LAMBDA Data Agent.
Agentic Workflows. The concept of an “agentic workflow” is not new. An agent senses the world or environment, and based on that information, the agent makes decisions, plans, and triggers the execution of a workflow of tasks and actions accordingly. The agent updates the state of the world and the results of the triggered workflows on a continuous basis. All this done autonomously without human intervention.
LLM-based agentic workflows. But within the domain and context of LLMs and LLM-based agents, the concept of an “agentic workflow” -although evolving quickly- is still in its infancy. Andrew Ng -the famous AI/ DL researcher and investor- recently popularised the concept of agentic workflows within the LLM context. A few months ago at Sequoia Capital, Andrew gave a very interesting talk on what's next for AI agentic workflows (video) and their potential to significantly propel AI advancements. And from a practical point of view, and again within the context of LLMs, checkout this blogpost on how to build an agentic workflow with CrewAI and Groq.
True agents in LLM agentic workflows. There are many AI researchers who say that LLM-based agentic workflows are very open-ended, and in essence not fully or truly agentic. These researchers argue that LLM-based agents: 1) don’t actually understand the world and can’t adapt to it continuously, 2) have no self-awareness of its existence 3) don’t have true reasoning & planning capabilities, as these are enabled by retrieving information from large datasets using pre-defined prompt templates. A few days ago, Druv -an expert in agent-based models- wrote on the concepts of agents, workflows and autonomy. This is a brilliant post: True agents in LLM agentic workflows: Lessons from Agent-based models for LLM-based agentic systems.
Hands-on, tutorials and practical guides
A new way to build scalable, modern agentic RAG pipelines with micro-services. RAG has massively evolved from the early days of naive RAG, to modular advanced RAG. Modern RAG systems now add an agentic framework on top of a retrieval system. This new framework handles reasoning, decision-making, and reflection on the retrieved data using micro-services. Blogpost: Build an Agentic RAG Pipeline with Llama 3.1 and NVIDIA NeMo Retriever NIMs
On AI Agents planning. Planning is a key aspect of intelligence. In AI/ ML research, planning has always been subject to many different debates and approaches. From a practical perspective and the point of view of LLM-based agents, this is a nice post on what planning means for an agent and how to improve it.
Building a SQL Agent. SQL is one of the most used languages in enterprise. There are quite a few startups building SQL agents but the challenges are to define reliable sub-agents at each state of a SQL pipeline, and to orchestrate them consistently. This is great video tutorial on how to build a reliable SQL agent with flow engineer and LangGraph.
New AI Agents, tools, platforms, and frameworks
The new Llama 3.1 450B’s agentic capabilities. This repo allows you to run Llama 3.1 as a system capable of performing "agentic" tasks like: 1) Breaking a task down and performing multi-step reasoning, 2) Using built-in tools like search or code interpreter, and 3) Zero-shot prompting: the model can learn to call tools using previously unseen, in-context tool definitions. Repo: Agentic components of the Llama 3.1 Stack APIs
A new multi-agent data system. This is a new open-source, code-free multi-agent data analysis system. LAMBDA uses two agents: the programmer agent and the inspector. The programmer generates code based on the user's instructions and domain-specific knowledge. The inspector agent debugs the code when necessary. Website and demo: LAMBDA: A Large Model-Based Data Agent.
Startups, Investors, and Enterprise
Why agents are the next frontier of generative AI - new MacKinsey report
Amelia - Conversational AI Agents to automate enterprise tasks
Mindset - Knowledge retrieval AI agents for companies
Outverse - AI Agents for instant customer support
Research Papers
A breakthrough in agentic web automation. This paper introduces Agent-E, a new agentic method for interacting with the web DOM (Document Object Model) and performing stateful navigations using hierarchical planning. Paper: Agent-E: From Autonomous Web Navigation to Foundational Design Principles in Agentic Systems.
Enhancing agentic learning up to human level. This paper introduces Discover, Verify, and Evolve (DiVE), a framework that discovers world dynamics from a small number of demonstrations, verifies the correctness of these dynamics, and evolves new, advanced dynamics tailored to the current situation. Paper: Enhancing Agent Learning through World Dynamics Modelling.
Introspection for agentic self-improvement. This paper introduces RISE, a new fine-tuning approach for agents self-improvement. The method enables models like Llama2, Llama3, and Mistral models to improve themselves. Paper: Recursive Introspection: Teaching Language Model Agents How to Self-Improve.
AI Agents London meetup
What an amazing meetup! We had two awesome talks on AutoGen and Camel AI agents frameworks. The speakers were great, and were also brilliant on answering lots of very interesting -and also challenging- questions from the audience. After the meetup, many of us went to the pub and had some really deep, thoughtful discussions on the reality, opportunities, and future of AI Agents.
Here are the slides from the meetup: AutoGen and Camel AI.
If you haven’t done it yet, join AI Agents London meetup and meet other like minded people interested in and building AI Agents.
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.
This discord link doesn't work :)