Your Next Training Script Might Be Written by an Agent

This session demonstrates using a coding agent (Codex) to orchestrate SFT of another model (Gemma 2B) on agent traces. The core bet: defining high-level constraints rather than scripts, letting the agent handle parameter sweeps, job scheduling, error handling, and evaluation.

This changes the workflow from manual MLOps to AI-assisted training pipeline generation. The agent integrates with Hugging Face Jobs for compute, TRL for training, and Tracker.io for metrics. It conducts smoke tests for safety before spending budget, then runs parameter sweeps across learning rate and sequence length.

The immediate implication is lower friction for domain-specific fine-tuning. A team could collect their own agent traces and train a small, focused model to emulate specific behaviors. The tradeoff is losing fine-grained control: the agent's script is verbose and decisions are guided by custom skills, which themselves require maintenance and versioning.

What's largely unacknowledged is that SFT only teaches imitation, not improvement. The resulting model cannot exceed the quality of the traces it learns from, and the evaluation focused on held-out loss rather than agentic capability. For production teams, this is a specialist distillation technique, not a path to general agent improvement.