Agent Harness Engineering
About the track
Most agent failures in production are from harness failures: the tools, permissions, loops, and validation around the model that nobody treats as a discipline yet.
This track is for the engineers running agents in production past the prototype, under real constraints like regulated data, legacy systems, and hard cost ceilings.
We’ll Cover:
- Tool permissions and sandboxes that survive regulated data
- Self-correction and feedback loops that catch errors before they reach users
- State and memory that stays under a cost and context-window ceiling
- Validation and evals that hold at production volume
- The agent loop itself: stopping conditions, retries, step and token budgets
Every talk comes from something the speaker built and measured, including the failure modes that forced a redesign. War stories and trade-offs. You’ll walk away with concrete harness patterns you can put into your own agents the same week.
Track host
Dave Scharbach
Executive Director, TMLS
See Track Lead Profile
Lorem Ipsum
Summit (2 days).
MLOps World | GenAI Summit 2026 is a two days of case studies, workshops, and expo on taking AI/ML and agentic systems into production – at the Etter-Harbin Alumni photo – full-bleed hero or browse files.