Beyond the Harness: Canvas for AI Agents

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. At the same time, there is untapped potential in the capacity of AI coding agents to create, self-correct and learn autonomously.

This talk draws on experiences from designing architecture governance models, applying them to design canvases for bounded AI autonomy. We will explore the intentional division of responsibilities between humans and agents, enabling a symbiotic relationship for safe innovation in a continuously changing ecosystem.

You'll leave with concrete techniques for:

1. Designing feedforward controls, feedback loops, and quality gates that work together as a coherent system
2. Applying multi-dimensional metrics to monitor and mitigate drift, decay, and runaway feedback loops
3. Structuring harness architecture using systems thinking, relying on adversarial pressure for continuous evolution