Summary
Microsoft engineers Ornella Bahidika and Joel Allou make the case that agent reliability is a control-flow problem, not a prompting problem. Drawing on their experience building Ace, a live AI voice tutor, they argue that asking an LLM to manage its own lesson state — tracking which step it is on, deciding when a student has mastered a concept, determining what comes next — is structurally unsound regardless of how much the prompt is tuned.
Their solution is what they call harness engineering: a state machine that encodes every lesson step (intro, teach, check, grade, advance, wrap), sends the model a tightly scoped contract for each step, validates the output, and advances state itself. The model — Claude Haiku 4.5 in their implementation — never decides where the lesson is; it only executes a specific, bounded action given a specific input. Live log recordings show the harness coordinating whiteboard drawing, queue clearing, section transitions, and lesson completion entirely outside the model.
The presenters offer a clear decision rule: if your agent’s behavior feels like a coin flip, move the decision logic out of the model and into code. They extend the pattern beyond voice to coding agents, ops runbooks, and onboarding flows — any scenario where deterministic, auditable control flow matters more than LLM autonomy. The cost is upfront engineering investment; the payoff is consistent, cost-effective, low-latency behavior at scale.
📺 Source: AI Engineer · Published July 20, 2026
🏷️ Format: Keynote Launch
