Marvin Prutlab, CEO of the emerging startup Convergence, and his co-founder Andy Tulis (CTO) met while working on recommendation systems and AI assistants at Shopify. The two joined enterprise AI platform Cohere and came up with the idea for the startup a few months ago. [Cohere] “Because we'd been using agents for a few years — three or four years,” Prutlab said. They realized the concept of agents was premature a few years ago.
The duo hired talent from Google DeepMind, Meta, OpenAI, and PolyAI, developed the product in just a few months, and attracted significant interest from investors.
Currently, most agents are designed for a specific workflow. Convergence's “proxy” agents tackle a wide variety of tasks with the idea that they can learn skills in the same way humans do by feeding them what is called “long-term memory.” This is done through what we call “large-scale meta-learning models” (LMLMs). They are trained to acquire skills that learn on their own.
Convergence has raised $12 million in a pre-seed round led by Balderton Capital, with Salesforce Ventures and Shopify Ventures also participating, which will be used to develop new models that will power the Proxy assistant.
“Few people have the experience and skill set of Marvin and Andy, which makes them perfectly positioned to tackle the complex technical challenges of a product like Proxy,” James Wise, a partner at Balderton Capital, said in a statement.
Human users are paired with proxy agents that can learn tasks and workflows, allowing the human worker to make more decisions than just the “grunt work.”
“If you look at the current landscape, a lot of companies are building these agents – for example, a sales agent, an HR agent, a financial operations agent, etc. But we're trying to take a different approach. We're trying to build the foundation for the first common class of agents, where users can define any type of agent they want that will do for them what they don't want to do at work as an individual,” Purtorab said. “We think this is a better approach, because we don't think that in the future, everyone is going to have 1,000 different little tools. Over time, things are going to consolidate, and that's where we come in.”
The strategy is to create consumer agents and then use them to inform how to train agents within enterprises. “Consumers are increasingly using it for online shopping, for things like groceries and recipes. On the enterprise side, they're using it for sales operations — entering and tracking a lot of information in Salesforce, for example — or for human resources operations, like tracking job applicants.”
So does that mean they're using consumers as guinea pigs for companies? “That's one way of saying it, because consumers have a lot of use cases that are simpler and have a broader variety, so it helps you get feedback faster, right?” Purtorab says.
But what do they think about the announcements Salesforce is making about agents? “I was at Dreamforce last week and I think it's the right direction. But I also think Salesforce is moving very far in the direction of very specific agents that are focused on specific tasks. At this point, we don't have the time to wait another six to twelve months until the model is a little better.”
Currently, the company is running a closed beta with testers, but will be releasing it to the public soon.