From Building to Directing: The Supervision Era Arrives

Pietro Schirano demonstrates that GPT-5.5 is qualitatively different: it can deconstruct an image into a harmonic sound and reconstruct it flawlessly, a creative task previously out of reach. This isn't just a toy—it signals that models can now execute complex, multi-step instructions without babysitting.

For engineering teams, the implication is a shift from hands-on building to directing agent swarms. Schirano’s demos—repurposing obsolete hardware, generating app icons instantly, and running Codex unsupervised on long tasks—show small teams can compress months into days. The bottleneck moves from coding speed to workflow design and instruction quality.

The tradeoffs are real: the demos are anecdotal, and reliability, consistency, and cost at scale remain unknown. Over-trusting unsupervised agents without governance will lead to chaos. As a founder, Schirano notes the new tension: the ability to build anything instantly collides with the other duties of leadership—fundraising, client relationships, team health.

The practical takeaway for engineering leaders is clear: start designing supervision-first workflows now. Hire and train for directing agents, not just writing code. The tools are coming; the discipline to manage them is not yet widespread.