Skill bundles

Measure current agent performance on the target workflow before RL intervention

Determine whether a workflow is actually amenable to RL improvement (clear rewards, sufficient volume, safe to explore)

Map an enterprise workflow end-to-end: inputs, decisions, tools, outputs, success criteria

Track and switch between reference model, current policy, and reward model versions during training

Configure vLLM or similar engines for RL workloads (batched generation, multiple completions)

Serve models at high throughput for RL rollout collection (not just user-facing latency)

Estimate and optimize GPU hours needed for RL training runs

Choose the best model checkpoint based on eval performance, not just training metrics

Diagnose and fix common RL training failures: reward collapse, mode collapse, KL explosion

Control how far the policy drifts from the reference model during training

Train reward models from human preference data, handle label noise and distribution shift

Tune learning rates, KL penalties, reward scaling, batch sizes for RL stability