Evidence from building AI systems in the real world.

Organized diffusion models execution. It turns ambiguous intent into an outcome with evidence by doing real work across agents, tools, context, workflows, and artifacts.

I ran a manager-worker agent team on 67 real tasks. A single session would have done the work better. The failure modes are structural.

Agent memory should be files in git, next to the work they describe, not rows in a vendor database.

A gate an agent waves itself through is not a gate. Quality gates for agent work have to be decisions against artifacts.

The unit of work is no longer a prompt. It is a loop with a driver, a body, exit conditions, and failure modes.

Production got cheap. The bottleneck moved: holding fragments together until the work is actually done.

Nine months of running Festival on real work, measured from the logs.