All blog posts
Every company will soon run on agents that learn from their own work: agents that take on real jobs inside the apps a business already uses, start on their own, and get better with every run. That is the future we’re building at Labelbox.
Today, we’re announcing more than $100 million in Series E financing, including equity and a credit facility, at a $1.25 billion pre-money valuation. The round is led by Infinitum Partners, with participation from Wipro Ventures, a16z, B Capital, Tau Ventures, Snowpoint Ventures, and others.
- $100M+Series E, led by Infinitum Partners
- $1.25Bpre-money valuation
- 400+customers build with Labelbox
- 90%+of leading US AI labs work with us
Data, environments, agents
Labelbox has been through three chapters, and the third just launched.
Data. Our first chapter was software for the people building AI: tools to capture expert judgment and turn it into training data. Many leading tech companies use our platform for enterprise-wide post-training programs.
Environments. Then came the ChatGPT moment, when transformers and reinforcement learning from human feedback made models broadly useful. The most capable models no longer learn only from examples. They get better by practicing real work. With Horizon, frontier labs train their models in reinforcement learning environments we build to mirror the full range of real work. We now work with 90% of the leading frontier AI labs and hyperscalers.
Agents. Models can now do real work, but assistants still wait for someone at the keyboard, and when people log off, the work waits. So we built agents that start on their own, work inside the tools a company already uses, and get better every run, for ourselves first. More than 50 now do real work across Labelbox, in every function from engineering to sales and finance. Last week, we opened that system to every team as Recursion Managed Agents, and we’re excited to bring it to the broad enterprise.
Toward systems that improve themselves
Horizon and Recursion serve different customers, but they work the same way. In Horizon, a model attempts a task in an environment, a calibrated reward says whether it succeeded, and the model improves. In Recursion, an agent finishes a job, a separate grader scores it against the team’s bar, and every graded run feeds the memory the next one starts from. Labelbox uses Recursion internally: more than 50 agents across engineering, research, product, sales, marketing, and finance start on their own from schedules, alerts, and tickets, and work around the clock. Every one is graded, and every one gets better.
The field calls this recursive self-improvement: systems that get better from their own work. We think of it in three parts. They reflect on their prior sessions and learn from them. They keep curated memory, so the next run starts from what the last one learned. And over time, they can be fine-tuned on their own best work. Each of those is a layer, and Recursion has all of them, from the sandboxes agents work in and the subagents and graders that check their work, to the memory and models that carry what they learn forward.
That’s deliberate. I believe the most important AI systems will be the ones that get better from their own work, with people setting the bar and software scaling it. Whether a customer is a frontier lab training its next model or a company running its first fleet of agents, we’re building every product to put that kind of improvement in their hands.
Innovating across layers
Systems that improve themselves take more than a good model, so we build every layer ourselves. We built our own agent harness from the ground up, designed for ultra-long-horizon agents. Our sandbox infrastructure runs millions of VMs, powering both Horizon and Recursion. We operate our own reinforcement learning post-training infrastructure, so we can post-train models quickly. And we run one of the largest expert networks in the world, more than 4 million people, which puts the best human creativity and intellect at our fingertips. Because we own each layer, we can co-design and optimize them as one system.
What comes next
Labelbox reached profitability this year while growing more than 5x year over year. With this financing, we plan to accelerate even further: more research and engineering behind Horizon and Recursion, and more of the people who bring both to our customers. Access to compute matters more with every chapter, and it will be central to this next one. Infinitum’s John Yetimoglu was an early investor in Groq and is now involved in building large-scale compute infrastructure at GroqCloud, and we’re glad to have that experience working with us.
Thank you to our hundreds of customers, from frontier AI labs and hyperscalers to startups and Fortune 500 companies, for holding us to the highest bar. Thank you to the Labelbox team for building through every reinvention, and to Infinitum Partners and all of our investors for their conviction.
A founder’s note
I came to America to learn to design and build aircraft and rockets, working on computational fluid dynamics and neural networks, and that’s where I met Brian. We spent our first years building hardware together, from wind energy systems to research labs for the International Space Station. Physical systems, software, and machine learning became the thread through my career.
At DroneDeploy and then Planet Labs, I saw how much of the opportunity in machine learning sat in data. The machines could collect enormous amounts of it, but as deep learning took off, the harder problem was teaching them what was worth paying attention to. It took people who knew what good looked like to turn it into something a model could learn from. In 2018, Brian, Dan, and I started Labelbox to help the people building AI do exactly that.
Data has remained the foundational challenge for model builders, and their biggest opportunity. Then reasoning models arrived, and reinforcement learning started to work in a general way: models could practice real tasks, be scored on the outcome, and get better. I believe that is the strongest force in technology today. I spent my career around machines that sense the world. These machines learn from it, and they let us build products and applications that improve themselves.
Labelbox is an American company, and there is a distinct, ruthless rationality to the American enterprise. It is uninterested in pedigree, popularity, or social niceties. It asks one question: does it work? We’ll keep earning our place the only way we know: by building things that work.