CJ Avilla and Harrison Stall use lessons from operating-system development to explain why a single coordinator can become a bottleneck. Their workshop demonstrates managed-agent flows that allocate isolated agents to phases such as scanning, review and synthesis, with examples spanning research, pull-request bug detection and explainer-video production.
The speakers distinguish an interactive coding workflow from an API surface that can be embedded in products or triggered on schedules. They show durable event inspection, budget pauses and continuation, and discuss how to evaluate accuracy and token costs rather than assume more agents automatically produce better results. A running operating-system experiment is explicitly unfinished, not a completed success.
The security section separates an agent's execution environment from its controlling harness and session state. CJ Avilla and Harrison Stall explain placeholder credentials replaced at an egress proxy, destination-scoped vaults, restricted network access, tool permissions and interruption. They also acknowledge current limitations: shared filesystem access is not inherently path-isolated, model-visible budgets are not the default, and the final discussion describes only one supported layer of nesting rather than unlimited recursive fan-out.
Watch on YouTube
