Harness Engineering: Moving From Hopeful Instructions to Enforced Control

Title: Harness Engineering: Moving From Hopeful Instructions to Enforced Control

Date: November 12, 2026

Speaker: Nicole Koenigstein

Registration Link

Your AI agent’s behavior depends on more than its model. Your harness determines what the model can see, which actions can execute, what state persists, and how the system verifies results and recovers from failure. What happens when those controls are missing?

This talk examines exactly that. It maps the core components of an agent harness, then makes their roles visible through experiments and animated execution walkthroughs. You’ll examine how removing controls changes outcomes, how oversized skill files affect output quality and cost, and how malicious skill instructions can cross trust boundaries when execution permissions are poorly enforced.

In addition, you’ll see how runtime skill compilation assembles relevant instructions and how harness controls govern their execution. You’ll leave with a practical understanding of which responsibilities belong in your model instructions, which require enforcement in the surrounding system, and why engineering your harness matters.

Additional Resources:
Harness Engineering (“Early Release” O’Reilly book free for ACM Members with Skills Bundle)
Expert Showcase: Architecture Meets Harness Engineering (free O’Reilly conference for ACM Members with Skills Bundle)
Harness Engineering for Browser Agents (free O’Reilly live course for ACM Members with Skills Bundle)

1 Like

In production AI-agent systems, which controls must be enforced outside the language model to guarantee that tool permissions, human-approval gates, and execution boundaries cannot be bypassed by prompt injection or malicious skill instructions?