Design a platform for long-horizon agent training environments

Agents are trained and evaluated on long tool-use tasks (up to about a thousand steps) inside sandboxed, versioned environments. Design the platform around the learning algorithm, not the algorithm itself.

Cover:

• how tasks, environments, allowed tools and reward definitions are versioned and published;
• running many isolated rollouts at once, recording every step, and resuming a rollout from a checkpoint you can trust;
• rewards that mix automatic checks with human review, and how you catch reward hacking and environment drift;
• comparing agent versions on fixed evaluation suites and deciding when to promote or roll back;
• isolation, privacy, recovery and audit.