Don't Let the LLM Drive - Ornella Bahidika & Joel Allou, Microsoft
Agent reliability is a control problem, not a prompting problem — harness the LLM, don't let it drive.
“Reliability was never a prompting problem. It's a control problem.”
Microsoft engineers built 'Ace', a live AI voice tutor, using a deterministic state machine harness that controls lesson flow rather than delegating control to the LLM itself. The core insight — that multi-step agent failures stem from control architecture, not prompt quality — led them to constrain each model call to a single 'neural contract' per state. This pattern enabled them to replace a heavy frontier model (Opus 4.7) with a much smaller one (Haiku 4.5), achieving reliability, speed, and cost efficiency simultaneously.