AI agent startup deals soar

Funding deals nearly doubled in the past year, AI2 releases new SOTA multimodal models for visual agents, and more

Author

Michael Cunningham
September 30, 2024

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

Media creators deal with a lot of difficulties, but finding out you’ve been an AI all along is definitely a new one. I assure you that Building AI Agents is written by a human…at least, I think I’m a human.

In today’s issue…

  • Funding deals for agent startups have surged

  • The original AI agent releases a new platform

  • SOTA multimodal models for visual agents

  • A framework to connect agent frameworks

…and more

🔍 SPOTLIGHT

Created using Dall-E 3

Anyone who closely follows the AI agent space—including yours truly—can tell that the number of startups launching in it has skyrocketed. Now, thanks to PitchBook, we can finally put numbers to this trend.

According to a new report by the venture capital and private equity database reported on by Fortune, 156 LLM startup deals have occurred in the past 12 months, up over 81% year-over-year. At the head of the pack of funders are VC stalwarts Y Combinator and Sequoia Capital, with 38 and 16 deals, respectively. The median value of these has also risen by nearly 47% to $3 million. Just as agents are a subfield of artificial intelligence that have come into their own with the explosion of interest in AI, we have seen subtypes of agent startups emerge—software engineering, customer support, sales, browser control, computer control, and many more.

The rise in venture funding tracks the broader growth of the AI agent field. Since this time a year ago, search interest in agents has increased nearly fourfold, according to Google trends, with a noticeable surge occurring as Salesforce made its aggressive pivot to being an “agent-first” company. This, however, was not an isolated catalyst, as other tech megacorps have simultaneously made bold pushes into AI agents—see the Agent Wars article below for just a few of them. Expect to see interest in and adoption of agentic tools—and with them, the amount of money pouring into the space—to continue to increase.

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

📰 NEWS

Source: TeamWave

Toran Bruce Richards and the team behind AutoGPT, the original AI agent which launched the field back in March 2023, have released a new low-code platform to allow developers, professionals, and businesspeople to build and deploy agents.

OpenAI has released a new model in its moderation API, seemingly named only “omni-moderation-latest”, which improves on the company’s prior models intended to check LLM outputs for harmful content such as violence and self-harm.

Open-source project MemGPT went viral last year for its ability to add long-term memory to commercial LLMs. Now, its creators have founded Letta to commercialize the technology, emerging from stealth mode at a $70 million valuation.

Cryptocurrency exchange Crypto.com introduced its AI Agent SDK, intended to allow LLM-based agents to take actions such as transferring cryptocurrency and creating wallets.

🛠️ USEFUL STUFF

Source: Allen Institute for AI

Molmo is a new family of open-source, multimodal LLMs by the Allen Institute for AI (AI2), which can take actions in response to visual stimuli and claim to outperform GPT-4o, Gemini 1.5, and Claude 3.5 on a benchmark of multimodal reasoning tasks.

AI agent and Web3 startup Fetch.ai is launching a lecture series on agents and multi-agent systems.

MotleyCrew is a new framework intended to enable agents from different frameworks to communicate with each other, allowing builders to benefit from the best features of each.

Zep Community Edition is the open-source version of the eponymous company’s agent memory framework. It creates a temporal knowledge graph from an agent’s interactions, serving as a long-term memory.

💡 ANALYSIS

Source: LinkedIn

Former OpenAI engineer and current Google AI Studio product lead Logan Kilpatrick appeared on Rowan Cheung’s podcast to discuss Google’s new Gemini models and their implications for AI agents.

This article addresses the question of why generative AI has not been as transformational as promised, arguing that it has been pigeonholed into automating narrow, isolated tasks when its true potential lies in more sophisticated workflows.

The authors of this article compare and contrast AI agents with previous generations of robotic process automation (RPA), claiming that LLMs now provide the contextual understanding necessary to tackle edge cases—and providing a useful map of the agent market.

This detailed and technical breakdown of the state of enterprise agents charts their movement away from large, closed-source models and towards smaller, open-source ones being run in-house, capable of taking actions in the real world and functioning together in multi-agent systems.

A roundup and ranking of 7 of the top 10 cloud providers’ agent capabilities.

🧪 RESEARCH

Source: arXiv

ChainBuddy is a new AI agent which allows users to describe a desired LLM workflow and assists them in instantiating it in LangGraph, making the latest “agent that builds agents”.

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