Trajectory integrated Daytona into its RL post-training stack through the platform’s Python SDK. While the process was seamless, Daytona’s leadership set up a shared Slack channel on day one to answer integration questions in real time.

With that foundation in place, Trajectory now runs its RL training loops on reliable, agent-native sandbox infrastructure. For every customer, Michael and his team use Daytona’s Snapshots feature to pre-configure sandboxes that replicate relevant tools, repositories, and workflows. In one deployment, they replicated a customer’s Salesforce instance, product documentation, and FAQs so agents could train on the same systems used in production.

These reusable templates eliminate repetitive setup work, unlocking faster sandbox provisioning. Combined with Daytona’s warm pool of snapshot-based sandboxes, startup times have dropped from 30 minutes to seconds. That speed enables Trajectory to run RL training loops at the pace their pipeline demands, accelerating model iteration and increasing training throughput.

Once provisioned, each sandbox runs for as long as a task requires. Throughout that window, it remains fully stateful, so modified files, actions, and context accumulated by the agent are preserved in real time. That continuity means each training cycle produces a more capable, accurate, and cost-efficient model than the last.

Critically, that performance scales. Daytona runs directly on bare metal with a proprietary scheduler, bypassing the overhead of traditional VM orchestration. This architecture enables Trajectory’s engineers to provision thousands of sandboxes in parallel, each one fully isolated to ensure reproducible results.

With Daytona’s integrated lifecycle controls, sandboxes are torn down automatically, freeing RAM, CPU, and storage. At the volume Trajectory operates, that precision keeps compute consumption tightly coupled to active training work. In turn, Michael can redeploy these resources toward new training runs.

Beyond the infrastructure itself, Daytona has become a strategic partner in Trajectory’s RL operations. Even after integration, the team remains closely involved, answering questions within minutes and conducting proactive check-ins. And because Daytona’s stack provides deep visibility into sandbox behavior, Trajectory’s engineers gain the insights they need to optimize performance and iterate across their RL pipelines.