Gap. The trainer agent autonomously picks models and runs Optuna 30-50 trials with no user say — start-training/schema.yaml only takes experiment_id, framework, and an agent-declared hyperparams dict.
Value. A real ML user wants to say "optimize PR-AUC, try only LightGBM + XGBoost, cap at 20 trials / 10 min." This turns an opaque autopilot into a steerable tool and directly bounds cost.
Suggestion. Add a training-config surface (a form or a structured prompt slot the orchestrator honors) for optimization metric, candidate model families, trial budget, and a wall-clock/cost cap.
Gap. The
traineragent autonomously picks models and runs Optuna 30-50 trials with no user say —start-training/schema.yamlonly takesexperiment_id,framework, and an agent-declaredhyperparamsdict.Value. A real ML user wants to say "optimize PR-AUC, try only LightGBM + XGBoost, cap at 20 trials / 10 min." This turns an opaque autopilot into a steerable tool and directly bounds cost.
Suggestion. Add a training-config surface (a form or a structured prompt slot the orchestrator honors) for optimization metric, candidate model families, trial budget, and a wall-clock/cost cap.