Agents fail in interesting ways. They don't crash, they drift. They don't error, they confidently do the wrong thing. Building a harness means designing for that specific kind of wrongness.
This talk draws on experience building harnesses for high-stakes AI workflows in life science environments, where regulatory compliance, audit trails, and failure containment are non-negotiable. These constraints turn out to be a gift: they force you to design control structures that any serious AI system needs.
You'll leave with concrete techniques for:
- Designing feedforward controls, feedback loops, and quality gates that work together as a coherent system
- Applying multi-dimensional metrics to agentic workflows: cognitive load, collaboration, architecture drift
- Structuring harness architecture using systems thinking with causal loop diagrams, leverage points, and evolutionary fitness criteria
Speaker
Nimisha Asthagiri
Technical Advisor @Thoughtworks & Co-Founder of BrainLoop AILabs, Previously Chief Architect @edX
Nimisha Asthagiri is a Global Director of Data and AI at Thoughtworks, a digital consultancy that integrates design, engineering and AI to transform enterprises to modern digital businesses. She leads digital transformations for her clients with strategic combinations of design thinking, change management, experimentation, and platform architecture. Her most recent focus is architecting agentic enterprises while applying systems thinking for responsible AI. Previously, she was Chief Architect at edX, driving intentional architecture for the next generation of large-scale online learning. Nimisha also serves as advisor and board member to emerging businesses, including serving as a consulting CTO. She began her career in Boston-based technology startups and holds multiple degrees from MIT. A seasoned technologist, Nimisha is passionate about fostering innovation through the amplification of diverse voices and the synergism of collective strength.
