In four accounts published on 26 October, AI produced reliable work where a team first wrote down how its own product works. Each team then checked the output against something the AI did not write. An investor's essay the same day argues for the same arrangement across a whole company: AI agents built by staff need rules on what they may touch, not a ban.
Blair Metcalf (27 Oct · 01), principal UX engineer at Reddit, the online community site, showed the podcast host Claire Vo how his team built Reddit Proto. The internal tool lets designers and product managers build working prototypes from Reddit's real components without learning Git, the system engineers use to manage code. More than 150 people in design and product use it, and about 700 prototypes exist. The team mapped every design token and component to production code, then refined that map from engineers' feedback on pull requests that passed or failed review. Code meant for production still goes through a generated requirements document and a draft pull request that engineers review.
Christine Vallaure (27 Oct · 02), a designer who runs design-system training, found that an AI agent turning a Figma file into code fills every unstated detail with a plausible guess. Focus order, error states and screen-reader text were invented, and none of it shows in a screenshot. She reversed the direction: the design system now lives in code, and the AI agent generates the Figma library from it. Each run also writes a file listing what Figma could not express and what the agent was unsure of, which she reviews by hand.
Arnaud Morvan (27 Oct · 03), a designer building the AI prototyping tool Axolo, tested a dozen small open AI models on about 2,500 questions drawn from real design systems. The models stalled at 60 to 65 percent correct on ambiguous cases until he gave each task a one-page guide written from his own design decisions. Three small models then reached 78 to 81 percent on cases the guides had never seen. This is a vendor's account of his own product, and one designer settled the hard cases.
The data platform at Netflix, the streaming service, sees roughly 100,000 failed data jobs on peak days. Its older rule-based diagnosis system could not classify about half of the failures it processed. Praneeth Yenugutala and five colleagues (27 Oct · 04) describe how they added an AI model without losing what the rules did well. The model returns only the first and last line numbers of an error, so the text shown always comes from the original log. The old hand-written rules serve as reference answers, and evaluation scores must pass set thresholds before any pull request merges.
Marc Andreessen (27 Oct · 05) wrote in the newsletter of the venture capital firm a16z on the day Replit launched a company workspace for AI agents. He argues that useful workplace technology arrives as chaos. Personal computers, spreadsheets and email each entered offices through employees before the company set rules, and early spreadsheets were often understood only by the person who made them. The companies that kept the benefit did not ban the tools; they added access rules, locked cells, version histories and audits. He argues that AI agents writing code are the next case, and that they need a governed place to work rather than a ban. This is an investor's essay praising a product launched that day, and it reports no team's measured results.
Each team gave its AI a written account of its own system: a component map, a code library, a one-page guide or a set of old rules. Each team also checked the output against something the AI did not produce, such as engineers' review, a list of gaps, hand-settled answers or frozen rules. Andreessen asks for the same at company scale: let employees build with AI agents, inside access rules and audits that make the work checkable. The inference, not any author's claim, is that the hard work now lies in writing and maintaining those references rather than in choosing an AI model.
Three of the four accounts come from people describing tools they built themselves, and none reports an independent measure of quality in production. Netflix calls its own main usage figure an observational comparison that does not show cause. Andreessen's essay praises a product launched the same day and rests on history rather than measured results.
Reddit built an internal prototyping tool on its real components and account data, and sends code meant for production through a generated requirements document and engineers' review.
Designer Christine Vallaure now keeps her design system in code and has an AI agent generate the Figma library from it, with every gap listed for her to review.
Designer Arnaud Morvan found that a one-page guide for each task lifted small AI models on design-system questions far more than switching to a different model did.
Netflix engineers added an AI model to their failure-diagnosis system, using old hand-written rules as reference answers and evaluation scores to gate every merge.
Marc Andreessen argues that companies should give the AI agents their employees build a governed place to work, as they once did for personal computers and spreadsheets.