When Augustine Mavor-Parker co-founded Vmax, he had one clear goal: enabling AI agents to learn autonomously and compound their capabilities through programmatic performance evaluation. This vision required training agents to generate, execute, and optimize their own tasks, a process that demands robust sandbox infrastructure to run reliably at scale.

At Vmax’s development velocity, concurrent provisioning of isolated and controlled sandboxes was non-negotiable. Their models generated tasks far faster than their system’s evaluation capacity, so generating one sandbox at a time would leave agents waiting for feedback and halt Vmax’s research. Because per-sandbox delays could multiply into significant bottlenecks at that scale, startup speed also had to be high to maximize throughput. 

To ensure accurate evaluation, each sandbox also had to include the dependencies and tools required by the task. Many of Vmax's tasks focus on software engineering, like learning to add new features to repositories. Custom development environments would ensure that outcomes accurately reflect task difficulty, advancing open-ended learning.

Augustine knew that building sandbox infrastructure in-house to meet these needs would require countless hours of spinning up fleets of cloud instances, managing uptime, and tuning startup times. Rather than pulling his lean research team away from core experiments, he set out to find a dedicated runtime platform that would provide the infrastructure and support to scale with Vmax's task pipeline.

Fortunately, Augustine didn’t have to search for long. After the AI research and development community pointed him to Daytona, he immediately knew that he’d found the solution to his dilemma.