Mercor’s agreement to acquire Deeptune marks the company’s second specialist purchase in 2026 and a clear move beyond expert recruitment. The company is building an integrated system in which people define skilled work, software recreates the tools where that work happens and verifiers measure whether an AI agent performed it correctly.
From expert marketplace to training infrastructure
Mercor announced on 9 July that it would acquire Deeptune, a New York company that builds simulated software environments for reinforcement learning. Financial terms were not disclosed. The deal arrived only months after Deeptune raised a $43 million Series A led by Andreessen Horowitz, with Mercor chief executive Brendan Foody participating as an angel investor.
Deeptune describes its product as a training gym for AI agents. A model can enter a safe replica of a spreadsheet, customer-relationship system or other enterprise application, attempt a task, receive a score and try again without touching live company systems. Mercor says the team has recreated hundreds of applications over two years.
“The constraint has shifted to the environments themselves: the places where models practice the work and get measured on whether they did it well.”
Brendan Foody, Mercor co-founder and chief executive
The second acquisition in a deliberate sequence
The Deeptune deal follows Mercor’s February acquisition of Sepal AI. According to legal adviser Orrick, Sepal specialised in high-quality training data, evaluation benchmarks and reinforcement-learning environments for frontier language models. Mercor said the team would strengthen its work on long-horizon research and the most complex tasks performed by people.
Together, the purchases suggest a consolidation strategy. Sepal adds specialised research and data-development capability. Deeptune adds the software layer where agents practise. Mercor contributes a large expert network and its APEX benchmarks, which score model performance across economically valuable workflows.
Why simulated work matters
Frontier models can perform strongly on exams while still failing inside ordinary business software. Real work depends on context spread across documents and applications, incomplete instructions and hundreds of small decisions that are difficult to represent in a static prompt.
Reinforcement-learning environments address that gap by giving agents repeated, measurable practice. Each environment needs the software where work occurs, tasks that reflect a professional objective and verifiers that can determine success. The quality of the task and verifier still depends heavily on human expertise.
“For enterprises deploying agents into their own workflows, it means one partner that can build the environment, staff the experts, and tell you whether the agent is ready.”
Mercor acquisition announcement
A more vertically integrated human-data market
Mercor says its network includes more than five million domain experts. That scale gives it a supply of people who can describe tasks, identify edge cases and judge whether an agent’s output would be acceptable in practice. Deeptune provides the simulated applications those tasks run inside.
If the combination works, Mercor can sell more than contributor hours. It can offer a complete development loop: map a workflow, build a safe replica, recruit the right professionals, create tasks and verifiers, train the agent and measure readiness. That moves the company closer to applied AI infrastructure and further from its origins as a matching marketplace.
The execution risk
Integration will determine whether the acquisitions become a coherent platform or a collection of specialised teams. Enterprise workflows vary widely, and a useful simulation must capture permissions, dependencies and failure states without becoming prohibitively expensive to maintain.
The strategy also concentrates responsibility. A provider that handles experts, identity, workflow replicas and evaluation data must maintain strong controls around confidential material and contributor information. Buyers will expect evidence that performance gains survive outside a simulated environment.
Mercor is betting that the next AI bottleneck is not access to people alone, but the infrastructure that converts professional experience into repeatable practice. Acquiring Sepal and Deeptune gives it more of that stack — and makes execution inside real workflows the standard it now has to meet.