Onboarding Daytona required only an API key and a single CLI parameter. After setting up both, Xiangyi had Daytona’s managed runtime platform ready to provision sandboxes at scale.

When Xiangyi runs SkillsBench today, BenchFlow automatically routes experiments through Daytona. Using Daytona’s Declarative Image Builder, his team specifies the exact base image, packages, and tools each task requires, from filesystem utilities to browser tooling. That way, each task starts from a consistent baseline without manual configuration.

Once provisioned, each sandbox enforces an isolated filesystem and process space, preventing agents from interfering with each other or destabilizing the local system. To add an extra layer of containment, Daytona enables Xiangyi and his team to create Docker-in-Docker environments by building pre-configured DinD images or manually installing Docker in a custom image. 

Because Daytona handles the full sandbox lifecycle, BenchFlow achieves these outcomes with minimal lift. State persists, so agents build on prior actions the way a task requires. That consistency holds as experiment volume fluctuates. Once trials finish and results are collected, Daytona tears down sandboxes automatically, maximizing resource utilization and eliminating unnecessary spend.

With this secure, elastic infrastructure, Xiangyi and his team can trust that benchmark results accurately reflect agent performance. As a result, researchers and engineers using SkillsBench can make evidence-backed decisions about where and how to deploy agent skills.

At the same time, Daytona empowers BenchFlow to host hackathons and competitions that require flexible sandbox capacity. During these events, sandbox demand can go from zero to tens of thousands of parallel containers over a few days. Daytona absorbs those spikes without additional overhead, keeping participants’ environments fully isolated from BenchFlow's core systems.

As BenchFlow’s reach grows, the lab now hosts wider initiatives across the AI community. For the BenchFlow Agent Skill Lift competition, each submission triggers 200 Daytona sandboxes to run its evaluation, a workload that scales automatically as entries come in. Running events at that scale consistently strengthens its standing as the standard for agent skills benchmarking.

With experiments running smoothly on Daytona, Xiangyi extended its sandbox infrastructure to BenchFlow’s agent learning pipeline. As his team iterates on the SkillsBench skill packages, Daytona’s pre-configured sandboxes ensure that any performance changes result from the skills themselves.

Beyond the platform itself, Daytona’s team has been a consistent resource for Xiangyi’s team. When questions arise, responses come the same day. And when BenchFlow needs additional sandbox capacity for a benchmark run or a hackathon, Daytona adjusts resource limits on request, ensuring no experiment is left waiting.