To advance agent autonomy, Martian launched ARES, one of the first open-source frameworks to enable online Reinforcement Learning (RL) for LLM coding agents. Rather than sticking to labeled solutions, ARES powers agentic advancement by enabling agents to run real coding tasks and receive real-time evaluations. 

However, ensuring task execution was both trustworthy and scalable came with strict runtime demands. “Agents are trained on a range of unique tasks, from SWE-Bench Verified to custom Harbor-packaged tasks,” shares Josh Greaves, ML Tech Lead at Martian. “As they work through each one, they edit files and run test suites in unpredictable ways.” To protect surrounding systems and keep experiments reliable, runs must be executed in tailored, isolated sandboxes.

But sandbox integrity was only half of the equation. With even routine tasks encompassing dozens of steps, attempting them one by one would take countless hours. Josh needed a way to provision hundreds of parallel sandboxes so agents could work through problems simultaneously and collect feedback for rapid, meaningful progress. Once a task was complete, each sandbox also had to be torn down to avoid resource waste and operational instability.

Josh knew from experience that standing up sandbox infrastructure at this scale required sophisticated container orchestration, resource scheduling, and lifecycle management. Rather than pulling engineers from research, he began looking for a managed runtime solution that could deliver the scalable infrastructure he needed out of the box.

It wasn’t long before his research network pointed him to Daytona. After seeing firsthand how their runtime platform handled hundreds of concurrent tasks for Terminal Bench, an industry-leading performance benchmark, Josh knew that he’d found the ideal partner.