100 million AI agents: Nvidia's CEO on the company of the future

Plus: LangChain's thesis on agent memory and the dawn of the agentic reasoning era

Author

Michael Cunningham
October 21, 2024

In partnership with

Welcome back to Building AI Agents, your biweekly guide to everything new in the AI agent field!

Apparently OpenAI’s new Swarm agent framework has gotten them in a bit of trademark dispute. I asked my lawyer what he thought of it but he told me that I was over my token limit

Our sponsor for this issue is Writer, whose platform gives you a simple interface to build AI tools powered by leading LLMs. Check them out below and show them some love

In today’s issue…

  • An online agent hackathon with $25,000 in prizes

  • The types of agent memory and how to implement them

  • Tips and tricks for agent developers

  • The agentic reasoning era begins

  • Automating agent workflow generation

…and more

🔍 SPOTLIGHT

Source: Wikimedia Commons

Every human will be a manager in the future, according to Nvidia CEO Jensen Huang, and you won’t have to haggle with your direct reports over sick days.

In a recent podcast interview, Huang laid out his vision of the future model of his company and likely many others: a hierarchical structure not unlike that of current organizations, but with AI agents forming the basic workforce, and humans acting as their managers. Nvidia’s current headcount sits at around 32,000, which Huang foresees as expanding to over 50,000 in the coming years—but these human workers will be dwarfed by armies of AI agents, numbering perhaps a hundred million. Regardless of whether this number represents a specific estimate or simply an attempt to convey the huge disparity in size between the biological and virtual workforce, the message is clear: in the company of the future, agents will be the rank-and-file.

Huang is not alone in envisioning an agentic enterprise. According to a recent Capgemini survey of 1,100 business executives, over half plan to begin using AI agents in the coming year, up from only 10% employing them now. Within the next 3, this number is expected to rise to 82%. Enterprises which neglect to augment their workforce will likely find themselves falling behind as their workers continue to spend time on menial tasks while those of their competitors are promoted to agent managers. This may lead to the birth of a new class of employee: the ultra-generalist, who uses their broad knowledge and skill base to identify the right problems to tackle, the execution of which is then delegated to AI agent workers.

In Huang’s prediction, this new regime will not lead to significant human job losses, as productivity will soar, generating ample demand for human overseers. But this state of affairs may not be a permanent one—with rote tasks outsourced to agents, the next class to find themselves under automation pressure will be the managers themselves. With efforts already underway to create AI systems that can build and provision new autonomous agents, humans will not remain the only agent builders and managers forever. While fully automated companies may seem like a far-fetched dream, AI capable of doing the work of an entire organization is the 5th and final step on OpenAI’s roadmap towards artificial general intelligence (AGI). With AI conquering task after task previously thought impossible for a computer, this future is not inconceivable. Regardless of whether this comes to pass, present-day executives would do well not to ignore the potentially transformative impact of an agentic workforce. The CEO of the world’s second most valuable company is not.

If you find Building AI Agents valuable, forward this email to a friend or colleague!

📰 NEWS

Source: TensorOps

TensorOps will host an online AI agent hackathon from November 14-17, sponsored by LangChain and LlamaIndex. A preparation event for participants will be held tomorrow night.

The computer maker released Lenovo AI Now, an LLM-powered agent running locally on the user’s machine, which can securely access their personal knowledge base to automate an array of tasks such as summarizing documents, configuring the computer, and generating content.

Presented by Writer

Writer RAG tool: build production-ready RAG apps in minutes

RAG in just a few lines of code? We’ve launched a predefined RAG tool on our developer platform, making it easy to bring your data into a Knowledge Graph and interact with it with AI. With a single API call, writer LLMs will intelligently call the RAG tool to chat with your data.

Integrated into Writer’s full-stack platform, it eliminates the need for complex vendor RAG setups, making it quick to build scalable, highly accurate AI workflows just by passing a graph ID of your data as a parameter to your RAG tool.

🛠️ USEFUL STUFF

Source: Flickr

Harrison Chase, CEO of LangChain, summarizes his and the company’s thoughts about implementing memory in AI agents, including different types of memory and how to endow LLMs with them.

A compilation of advice on developing agentic apps from a variety of players in the field such as Replit CTO Luis Héctor Chávez.

A video tutorial on how to LangGraph to create an agent that runs without any external API calls using an open-source model running on Ollama.

This new package promises to overcome the transparency limitations of existing frameworks such as LangGraph and AutoGen by continuously adding to and reading from a single memory “tape” in a manner reminiscent of a Turing machine.

💡 ANALYSIS

Source: Wikipedia

This piece by tech VC powerhouse Sequoia Capital explores the implications of a new epoch, ushered in by OpenAI’s o1 model, in which agents’ capabilities are constrained only by reasoning time, and the switch to a new “service-as-a-software” model for agent companies.

In this episode of Y Combinator’s Lightcone podcast, Replit’s CEO Amjad Masad discusses the power of the company’s software engineering agent and the future promise and challenges of AI agents.

Software development agents have had a rocky path to usefulness, and challenges remain, but a new report projects that 80% of software engineers will soon have to reskill as an increasing fraction of code is written by AI.

The author of this piece points to supply chain company Flexport as a successful case study in using generative AI to achieve substantial savings through automation.

🧪 RESEARCH

Source: arXiv

A paper by the team behind MetaGPT, one of the first AI agents, proposing a new method of representing agent workflow optimization as a search problem to be tackled using Monte Carlo tree. The resulting workflows significantly improve over manually engineered methods at a fraction of the cost.

Thanks for reading! Until next time, keep learning and building!

If you have any specific feedback, just reply to this email—we’d love to hear from you

Follow us on X (Twitter), LinkedIn, and Instagram