Judgment

I ended the last post with a question: the next challenge is not building the Judgment Log. It is whether anyone writes in it once the deadline is two hours away. That question only has a useful answer if the artifact is light enough to actually use. So instead of arguing for it further, I want to show it. Take a fictional but familiar scenario. A checkout flow. A promotional window. A promo code validation feature built using AI-assisted development. It shipped. Three weeks later, it broke when the campaign introduced expired codes. In the post-mortem, nobody could answer the three questions that mattered: what did the PM cut and why, what did the designer choose between, and what did the engineer override. ...

In April, Meta employees burned through 73.7 trillion tokens in roughly thirty days. The company found out not because spending crossed some alarming threshold, but because an internal leaderboard, nicknamed Claudeonomics, had turned token consumption into a competition. Employees and teams were ranked by how much they used. The system did exactly what it was built to do: usage went up. What it could never show anyone was whether any of that usage produced something worth the cost. Meta is now dismantling the leaderboard in favor of a centralized monitoring platform called AI Gateway, built to track spending in real time and flag unusual spikes. ...

In 2012, I was given a project at Dalet that I did not fully understand at the time: move the entire company from waterfall to Agile. We were a mid-size product company with engineering teams spread across time zones, and waterfall was doing what waterfall always does at that scale, creating the illusion of control while real problems accumulated in the gaps between phases. The transition worked. And the productivity gains were not incremental. Engineers who had spent months waiting for approval to write code were suddenly shipping every two weeks. The business could see working software, give real feedback, and change direction without blowing up a six-month plan. It felt like the industry had discovered something fundamental. ...

In a post I wrote last December, I discussed the production triangle: the constraint that governs every production system, including software. You can optimize for two of three dimensions, time, quantity, and quality, but never all three. Push velocity while adding features, and quality absorbs the cost. Every time, without exception. This constraint is not new. Kahneman mapped the cognitive mechanism in Thinking, Fast and Slow: under time pressure, we disengage System 2, the slow, deliberate reasoning that catches the things that will hurt us later, and default to fast, intuitive pattern-matching. Kuutila et al.’s 2020 systematic review of the software engineering literature confirmed the outcome: increased throughput, decreased quality, and a root cause that almost always traces back to the pressure itself being a product of bad estimation. The triangle is not a theory. It is physics. ...

Gartner published a prediction this week: by 2027, 40% of enterprises will demote or decommission autonomous AI agents due to governance gaps that are only identified after production incidents. The root cause they name at Level 3 of their autonomy framework is approval fatigue. Agents execute actions, writing data, sending communications, and modifying configurations, but only after explicit human approval. Under time pressure, that approval becomes reflexive. The control degrades, and the risk compounds in silence. ...

The bar for building has collapsed. Anyone with an AI assistant can produce a prototype, a draft, a working script. That is not the challenge anymore. The challenge is identifying people who can tell whether what was built is right, catch what is wrong, and make the call on what to do about it. That is a different hire than the one most managers have been trained to make. And the stakes are higher now, because your hiring decision is no longer a baseline. It is a multiplier. ...

Your team closed 47 tickets last sprint. Deployment frequency is up. Lead time is down. The sprint review looked clean. And you have no idea whether your team is getting better at what matters. Every metric on that dashboard measures motion. None of them measures whether the team made the right decision. In an era where AI generates the first draft of almost everything, code, architecture proposals, incident summaries, and product specs, that gap is no longer academic. It is the difference between a team that compounds in judgment and a team that accelerates toward the wrong destination. ...

For decades, the path was obvious. A junior engineer joined a team, got handed a bug nobody else wanted, fixed it, broke something else, fixed that too, and over a few years accumulated the scar tissue that turned into judgment. Senior engineers were not trained; they were grown, slowly, by the system itself. That system has stopped working at both ends. At the entry point, juniors aren’t being hired. AI makes a senior dramatically more productive and a junior only marginally so, which makes the rational move, for any single team, in any single quarter, to skip the junior hire and let an experienced engineer with AI do the work two juniors used to do. Industry-wide, the pipeline is being shut off before it starts. ...

The Lesson That Used to Be Unavoidable Every senior engineer I respect has a similar war story. They wrote a Python script that was too slow, and someone told them to learn what a list comprehension actually does under the hood. They built a React app that re-rendered itself into a coma, and they had to crawl back into the DOM to figure out why. They shipped a service that fell over the first time real traffic hit it, and they spent a weekend learning what a connection pool is. ...

Working with AI to generate code is extremely satisfying. In a matter of minutes, you get something that looks great and, in most cases, does what you wanted it to do and even more. But many times, what looks ready for production is far from being production-safe. A large-scale study conducted by two researchers at FernUniversität in Hagen analyzed 7,703 files from public GitHub repositories explicitly attributed to AI tools. Using CodeQL, the researchers identified 4,241 CWE instances across 77 different vulnerability types. While 87.9% of the analyzed AI-generated code contained no identifiable CWE-mapped vulnerabilities, the risk came from code that appeared to work fine. It compiled, it solved the visible task, but it still carried hidden assumptions, unsafe patterns, and security debt. ...