- Multi-agent systems are distributed systems of specialized AI agents working together toward a collective goal.
- VCs at NEA, Mayfield Fund, and Foundation Capital are eager to invest in this latest technological development in AI.
- Despite the nascent space, startups like Cognition and Sakana AI have each raised over $100 million.
The future alluded to by Spike Jonze's 2013 film "Her" is becoming closer to reality. Startups are building AI agents that collaborate to fill out forms, dial customer service reps, and order food delivery. These "multi-agent systems" are the next hot thing, according to investors from NEA, Mayfield Fund, and Foundation Capital.
Multi-agent systems are distributed systems of multiple specialized AI agents that work together toward a collective goal. These agents break down a task into multiple smaller steps, each tackling a specific task to achieve a broader objective.
"You give it a goal, and it breaks down that goal into a series of steps into a sequence. It can then execute that sequence," said Foundation Capital's Ashu Garg.
Take, for example, planning a trip to Italy, said Garg. A large language model may do a "decent job of [presenting] an itinerary, but that's not really planning a vacation." LLMs, like OpenAI's GPT-4, respond in a question-and-answer format and may lack the reasoning power to execute all the steps involved in planning a trip.
In contrast, multi-agent systems are built on top of LLMs and are designed to execute a task from start to finish. In the Italy example, agentic systems promise to plan the entire trip: booking flights and hotels, making restaurant reservations, and recommending activities. Throughout the process, the agents aim to continuously validate their work for accuracy and completeness.
The validation step is critical for multi-agent systems, said NEA's Madison Faulkner. By defining success criteria, these systems can autonomously verify whether they meet their goals throughout the
process. "That's where you unlock this very powerful, scalable approach to building agentic systems."
Additionally, AI agents collaborate throughout the process, whether by calling APIs or accessing relevant data. When necessary, these systems may also involve human input, commonly referred to as human-in-the-loop systems.
A crop of startups is already building in the nascent space. Many focus on specific use cases at the application layer that general-purpose LLMs don't adequately address. For instance, Regie AI, which raised $20.8 million according to PitchBook, uses "auto-pilot sales agents" to automatically source leads, draft customized emails, and follow up with buyers. Six months after it launched, Cognition raised $175 million at a $2 billion valuation, according to PitchBook. The company is behind Devin, an autonomous AI engineer that can execute complex engineering tasks.
Agentic startups can "come up with solutions to quite niche, or more defined problem sets that these larger AI companies may not be able to tackle," said Warren Hui of Soul Ventures.
Others focus on the infrastructure layer to power multi-agent systems. Emergence, which raised about $100 million, orchestrates between first-party and third-party models to automate enterprise knowledge work, such as summarizing reports or processing claims.
"Our company is trying to bring to the table this deep idea of what it means [for agents] to self-improve," said Satya Nitta, CEO of Emergence. The goal is for "an agent can control and operate all kinds of software systems or compute systems the way a human can."
Another startup, Phidata, transforms LLMs into AI assistants by enabling real-time data access, such as searching the web or querying a database. "Our inbound is through the roof," said Ashpreet Bedi, CEO of Phidata. "We haven't been pushed back on price even a single time."
Meanwhile, some straddle both the LLM and agentic layers. One example is Sakana AI, a Japan-based startup that raised $155 million, according to PitchBook. Sakana AI is a massive foundation model with elements of a multi-agent system, said Faulkner. The large-scale model breaks down its expertise into smaller specialized models, focusing on different domains like Japanese and math.
Another example is Hippocratic AI, a foundation model focused on healthcare that raised $120 million, according to PitchBook. It has partnered with Nvidia to build voice-based AI agents for patient-facing tasks, according to the company's website.
With numerous startups entering the space and investors pouring in money, the vision of "Her" seems closer than ever. "For the first time in history, you can actually let an agent loose to automate multiple workflows," said Nitta.
Mayfield Fund's Navin Chaddha is optimistic — AI agents will serve as teammates to workers. "I assume somebody like me is going to be working with 50 to 100 agents… I'll have a chat agent, support agent, service agent, research agent. It's just endless," said Chaddha. "And then these agents will have to talk to each other too."
Correction: August 1, 2024 — An earlier version of this story misstated the total funding the startup Emergence has raised. It is about $100 million.