Convergence AI has only existed since June 2024 but already it’s raised one of Europe’s largest pre-seed rounds ever: a $12m deal led by Balderton Capital, announced last September.
The London-based startup has developed a prototype AI agent named Proxy which it claims is — in some respects — outperforming agents tested by leading AI companies like OpenAI and Anthropic.
AI agent makers say their tools can accomplish valuable real-world tasks, rather than merely provide assistance to humans. If they’re right, companies will eventually be able to “hire” autonomous agents as full-fledged workers. Investors are betting on an agent-run future — Europe’s agent developers fetched just over €1bn last year.
With a team of hires from Google DeepMind, Meta, OpenAI and PolyAI, Convergence AI’s coming to market with a strong hand. We asked CEO and cofounder Marvin Purtorab to break down his business strategy and explain how the agent sector will develop.
B2B or B2C?
Convergence is still figuring out what its agents will be used for.
The basic idea is that Proxy should be able to master a range of tasks and work across industries, acquiring skills like a human would through long-term memory and continual learning. “In the long-run, the aim is to build anything that can get arbitrary work done,” Purtorab says.
Convergence is experimenting with its target audience. “The AI agent space is just over a year old. The reason why we went [to the] consumer [with our agent] first is because it helps us gather a lot of information on what users want to do with it.”
Purtorab wants customers to run amok and then use their feedback to inform how to train business-facing agents. “People will probably prefer to automate things at their job [rather than] automating things in their personal life.”
Pricing AI agents
As with any new piece of tech, pricing will depend on client focus. For now, the most common agent revenue stream is a pay-per-seat and/or subscription model.
“Some people are trialling pay-per-solved-tasks, which is attractive because agents aren’t that powerful yet,” Purtorab explains. “But they’re getting more powerful every week, if not day, and we’re going to have very general agents in a year or two.”
At that point, “agents will probably settle down into a subscription model”. He says it’ll be like how companies pay salaries, which in any case are “basically subscriptions to human labour.”
There are already tons of agent companies, “but most of them are vertically-focused on one use case at a time,” Purtorab says.
Convergence wants to develop “general” jack-of-all-trade agents. “The big hurdles [to get here] are reasoning and orchestration,” Purtorab says. General agents will need to be savvy at navigating websites, coding, making phone calls and interacting with other robots.
“We also want to give everyone the ability to build their own agents — not just the specialised AI firms. This should be so simple that everyone can do it.”
Hallucinations
Trust in agents will depend on how much “hallucinating” (i.e. waffling or lying) they do. “We ask Proxy to check in with users a lot. Where you don’t fully trust agents to do something, you have to agree [to their actions].”
Some hallucination is probably fine if it’s comparable to mistakes a human would make, Purtorab says.
Agents will be trained to keep tabs on each other, he adds. “Each deliverable is essentially double checked by another and if there’s something that seems off, there’s a much higher chance it’ll be called out this way.’
A threat to jobs?
Right now, agents aren’t much of a threat to humans — but AI founders say they could start to substitute for human workers in some jobs.
Things could get blurry. We’re traveling to a future where computers are using computers to carry out tasks; where robots are doing business with other robots.
“There’s initial resistance to any technology. But a product gets adopted if it makes people’s lives better, not worse,” says Purtorab.
“When automation keeps growing, we probably need to tax it quite strongly — it’s natural. I don’t think there’s anything bad about that. [But] if we get to a point where AI takes whole jobs, I’m not sure what the social contract will be.”
Read the orginal article: https://sifted.eu/articles/can-this-london-based-ai-agent-maker-beat-openai-and-anthropic/
