Marvin Purtorab, CEO of new startup Convergence, and co-founder Andy Toulis (CTO) met while working on a recommender system and AI assistant at Shopify. Both joined enterprise AI platform Cohere before coming up with the idea for the startup just a few months ago. “Honestly, we left (Kohia) because…we had been playing with agents for several years, maybe the last three or four years,” Partlove said. They realized years ago that the concept of agents was premature.
Hired from Google DeepMind, Meta, OpenAI, and PolyAI, the duo built the product in just a few months and received significant interest from investors.
Most agents today are designed for specific workflows. Convergence’s “proxy” agents work across many tasks with the idea that they acquire skills in the same way as humans, by giving them so-called “long-term memory.” This is done through what some call a “large meta-learning model” (LMLM). They are trained to acquire self-learning skills.
Convergence has raised $12 million in a pre-seed round led by Balderton Capital. Salesforce Ventures and Shopify Ventures also participated in this round, which will be deployed to develop new models to power Proxy Assistant.
James Wise, partner at Balderton Capital, said in a statement: “There are few people with the experience and skills that Marvin and Andy have, making them well suited to tackle the complex technical challenges of a product like Proxy.”
Human users are paired with proxy agents that can learn tasks and workflows, allowing human workers to make more decisions than “grunt work.”
Purtorab said: “If you look at the current landscape, there are a lot of companies that are building ‘narrow’ agency-like services such as sales agencies, human resources agencies, financial operations agencies, etc. Our view is to take a different approach. We’re trying to lay the foundations for the first general class of agents that can be categorized into any type of agent you want, depending on your user, to do things for you that you don’t want to do personally or professionally. I’m here. do. We think that’s a better approach. Because I don’t really see a future world where everyone has 1,000 of these different little tools. As time goes on, things will become more integrated and I think that will be our strength. ”
The strategy is to create a consumer agent and use it to inform how to train agents within your enterprise. “Consumers are increasingly shopping online for things like groceries and recipes. Businesses are using it for sales activities, such as entering a lot of information into Salesforce and tracking it. Or talent tasks like tracking job applicants.”
So are they using consumers as a kind of guinea pig for companies? “That’s one way of saying it. Consumers have a lot of simpler, broader types of use cases, and that’s where feedback comes in.” It helps you get it more quickly, right?” Pultlove said.
That being said, what do they think about the kinds of announcements Salesforce is making to agents? “I joined Dreamforce last week and I think it’s the right direction. But , I also think they’re moving further into something like a very narrow agent focused on one specific task… At this point, the time to wait another 6-12 months for the model to develop a little bit is No better.
Currently, the company is running a closed beta with testers, but it is expected to be released to the public soon.
