Workshop: Training Agents at the frontier of AI Industry
Enterprise customers training agents prioritize latency, accuracy, and de-risking over cost or model novelty
“Ultimately, what they care about is they have an IT budget that maps to some strategic priority, and then they want to improve on some business metric that actually helps them justify that spend.”
A Microsoft research engineer shared practical learnings from helping traditional enterprises (manufacturing, energy, industrials) train AI agents using RL environments built from production traces. Key insight: enterprise priorities are latency, accuracy, and containment — not cost savings or chasing frontier models — and agents often have 100+ tools requiring classic distributed systems optimization. The talk outlined a pipeline from OTel-instrumented traces to RL environments with synthetic perturbations and verifiable reward signals.