https://blog.mornati.net/tags/code-review/Recent content in Code-Review on Mornati BlogHugo -- gohugo.ioen-US© 2026 Marco MornatiTue, 03 Nov 2026 18:00:00 +0000https://blog.mornati.net/one-year-later-a-stronger-model-reviews-my-vibe-coded-app/Tue, 03 Nov 2026 18:00:00 +0000https://blog.mornati.net/one-year-later-a-stronger-model-reviews-my-vibe-coded-app/I built Cyber Code Academy in December 2025 with Cursor and Copilot, and had the AI of the time check its security. One year later, a single prompt to Claude Opus 5.5 found 50 problems, two of them critical. Anyone could delete the challenges, and any learner could fake their own test results. The bugs, the fixes, the redesign, and a second prompt that showed the exercises were unfair too.https://blog.mornati.net/three-agents-at-95-percent-dont-give-you-99-99/Tue, 20 Oct 2026 17:00:00 +0000https://blog.mornati.net/three-agents-at-95-percent-dont-give-you-99-99/Three agents at 95% each should give 99.9875%, if they fail independently and if you can tell which answer is right. Neither holds. With a voting step and realistic task difficulty the same three agents land near 96%. Why, what the data says about mixing model families and reviewers, and what we really know about human defect rates since the 1980s versus models from October 2025 and October 2026.https://blog.mornati.net/six-weeks-in-one-night-building-a-product-in-the-agentic-era/Mon, 19 Oct 2026 16:00:00 +0000https://blog.mornati.net/six-weeks-in-one-night-building-a-product-in-the-agentic-era/I gave a human development plan to a team of agents, an Opus spec writer, a Sonnet developer and an Opus reviewer, and let them build the first phase of Graft overnight. Six planned weeks became one night, at about 0.5% of the human cost. But the bottleneck didn’t disappear. It moved to us.https://blog.mornati.net/you-dont-need-scion-graft-review-layer-for-agent-pull-requests/Sun, 18 Oct 2026 18:00:00 +0000https://blog.mornati.net/you-dont-need-scion-graft-review-layer-for-agent-pull-requests/Anthropic’s Scion solves a storage problem only a few companies have. The workflow problem behind it, too many agent pull requests for the humans who review them, is everyone’s. Graft, now public, puts a checked intent card, a risk class no model can lower and a policy verdict on every pull request, so reviewers read the card before the diff. What ships in v0.0.1, how a reviewer’s day changes, the video, and what the live test checks.