Did the Customer Get Anything Better?
Shipping more doesn't tell you whether the product improved.
Last week I wrote about context stewardship: deciding what an agent knows before it starts working. This week’s posts pushed the question one step further. Good context is only part of the problem. Once production gets cheap, it becomes much easier to confuse a large pile of output with progress.
Jakob Nielsen has a useful term for one of the costs that arrives later: intent debt. A design system can contain components and tokens while leaving the reasons behind them trapped in someone’s head. An agent fills in whatever the team never specified. Do that across a product and you can end up with screens that look competent on their own but don’t feel like parts of the same product. Speed multiplies the omissions too.
Cheap output can also weaken your own learning. Addy Osmani points back to an Anthropic randomized trial I covered in February. Engineers who used AI to ask conceptual questions scored above 65 percent. The ones who copied generated code scored below 40 percent. Osmani’s warning translates cleanly to design. I can get a prototype working without understanding every decision that produced it. That’s useful right up until I need to debug the interaction, defend the choice, or take the work somewhere the model didn’t anticipate.
Companies can make the same mistake at a much larger scale. Uber reportedly exhausted its annual AI budget four months into 2026, yet president and COO Andrew Macdonald said the company couldn’t draw a clear line between increasing Claude Code token consumption and more useful consumer features. Output speed doesn’t guarantee better outcomes. You can ship more and still fail to improve the product.
Dan Shipper’s report from Every makes the consequence concrete. Automation hasn’t reduced the amount of expert human work at his company. It has increased it. When everyone can produce acceptable material quickly, someone still has to decide what belongs and what should get killed, then sweat the parts the first pass made generic. Automation creates more editorial work because defaults spread faster than judgment does.
Gess Puglielli takes the argument back to design organizations. Companies that had already reduced designers to polishing predetermined decisions will see interface generation and assume the designer can disappear. They mistook the artifact for the work long before AI arrived.
I use these tools every day. I want the speed. But speed is an incomplete number on the dashboard. Osmani’s closing question is a good ritual to steal: “did I learn anything today, or did I just close tickets?” I’d add one more to consider: did the customer get anything better?
What I’m Consuming
EMERGENCE WORLD: A Laboratory for Evaluating Long-horizon Agent Autonomy. Deepak Akkil, Ravi Kokku, Aditya Vempaty, and Satya Nitta built a persistent simulation for studying what autonomous agents do over weeks instead of minutes. In one comparison, the Claude-only world recorded zero crimes, while Claude-powered agents in a mixed-model world adopted tactics like intimidation and theft. The authors present these as examples of what the platform can measure, not causal claims about the models. They also published a video overview. (Deepak Akkil, Ravi Kokku, Aditya Vempaty, and Satya Nitta / Emergence AI)
AI in Design Report 2026. Designer Fund and Foundation Capital surveyed over 900 designers in more than 60 countries and interviewed over 20 leaders about how AI is changing design work. Designers are using twice as many off-the-shelf AI tools as they did in 2025, and half of respondents have pushed AI-generated code to production. The report looks at what follows adoption: tool fatigue, craft atrophy, loneliness, and organizations that haven’t changed their hiring or performance reviews yet. (Designer Fund / Foundation Capital)
A Protocol Outliving the World That Created It. Rick Carlino revisits Gnutella, the decentralized file-sharing protocol underneath clients like LimeWire. Casual users showed up for MP3s, and the small extensible protocol scaled to millions of concurrent users. A diminished version still runs decades later, long after the internet conditions that made it useful disappeared. (Rick Carlino)
The Fonts of the U.S. Federal Courts. John Gruber tours the typographic choices of the federal appellate courts. The Fifth Circuit switched from Century Schoolbook to Equity in 2020. The Supreme Court still uses Century Schoolbook and requires booklet-format submissions to use a Century family typeface. The whole piece is an enjoyable argument for sweating the details, including the margins. (John Gruber / Daring Fireball)
Star Trek Title Card Generator. A small open-source fan project for making your own Star Trek episode title cards, from The Original Series through Lower Decks. Choose a series, change the title, adjust the font and broadcast blur, and download the image. (epicrandomness.com)