Training AI agents takes more than fast token generation.
Tools run. Tests stall. Training batches wait.
We cut fixed-weight batch collection time by 66.9% on a Terminal-Bench-based workload by improving routing, sandbox execution and scheduling.
Here’s how: