The Code Governance Layer Your AI Stack Is Missing

Qodo CEO and co-founder Itamar Friedman on why the quality model built for human-authored software does not survive AI velocity, and what replaces it. The talk runs from generative adversarial networks in 2016 to a live look at cross-repo context across 5,000 repositories, with two audience polls along the way.

Topics covered:

The old quality model ran on human judgment, and that judgment left the loop. Standards and PR review worked because the team writing the code lived with it. METR found developers felt roughly 20% faster with AI while measuring 19% slower. Velocity rises; the decisions underneath move the other way.

Generation needs an adversarial counterpart. The framing goes back to GANs: a generator paired with a discriminator, each making the other better. The system that writes code should not be the system that verifies it, for the same reason financial audits and pen-tests stay independent.

Quality is a per-system measure, not a per-repo one. One company described itself as a monorepo shop and turned out to have 800 repos. Cloning five locally is not a context strategy. The demo maps 5,000 repos and more than 10,000 connections between them, generated automatically rather than documented by hand. Only 19% of attendees had that view of their own code.

Governance is a gateway, not just a gatekeeper. Adding AI review after AI code generation is necessary and not sufficient. The layer spans planning, writing, pre-PR, merge, and post-merge decision records. Aviation and autonomous driving scaled governance in step with autonomy. Agentic development has not.