Review Debt and What It Costs

00:00:04 Sachin Gupta at eBay put together something new this week. Not a model, not a tool — a metric for measuring the thing every engineering team is feeling but nobody has a number for yet. He calls it review debt. It's the gap between code your agent produced and code humans have actually reviewed, trusted, and understood.

00:00:26 He's not saying coding agents are bad. He's saying they're creating a kind of interest that compounds daily, and nobody's counting it. His evidence is pretty stark. GitHub's 2025 October report — which covers almost every public pull request on the planet — shows commits climbed twenty-five percent year over year, while review comments dropped twenty-seven percent in the same period.

00:00:52 Code production went up. Review attention went down. They moved in opposite directions. Faros AI tracked teams along the AI adoption curve in their 2026 benchmark. Median PR review time is up four hundred forty-one point five percent. That means reviewed PRs take five and a half times longer than they used to.

00:01:14 Thirty-one percent more pull requests are now merged with no review at all. Every one of these numbers is real, Gupta notes. But the usual dashboard metric — PR throughput up eight percent, AI usage rose sixty-five percent — is generous in ways that don't tell you what you actually need to know.

00:01:34 PR count goes up when one PR splits into seven. Median PR size going up is bloating, not benefit. Cycle time goes down when reviewers stop pushing back. These metrics tell you the speed of production, not the speed of trust. Gupta's framework runs five signal families across ten deterministic checks — no large language model acting as judge, because those scores become a moving target and you can't put them on a slide for leadership.

00:02:04 The checks are: developer size and coupling, test evidence gaps, directory and ownership spread, AI authorship indicators, and evidence and rationale gaps. Each PR gets scored zero to one hundred and sorted into bands from low burden to high. What makes the framework useful isn't the score itself — it's that it names the compounding loop.

00:02:28 Agents learn from your code base, which includes unreviewed code. Reviewers shift from architectural decisions down to syntax checks. Leadership resets velocity expectations without hiring review capacity in proportion. Each loop on its own is survivable. Together you get a runway, and then you're just counting how long until the debt calls.

00:02:51 Whether any team that hasn't run these checks backwards over their last two hundred merged PRs can actually say they know what they're accumulating remains an open one.

00:03:03 There's another number hiding in this week's data. TypeScript passed Python as GitHub's most popular language last August. Roberto Stagi, CTO of Reto — a context layer for AI agents — walked through the move yesterday and pointed to the same driver everyone else is seeing but treating differently: coding agents.

00:03:26 The GitHub report used the same lens here as it did in 2024 when AI led Python to the top. In 2025, AI leads TypeScript. The mechanism is a feedback loop Stagi describes plainly. Coding agents like Cursor, Codex, and Cloud Code became the default way to build applications.

00:03:46 They default to TypeScript. New apps ship TypeScript. More TypeScript in the wild means the next generation of agents trains on more examples. They get better at it, so more people choose it. Anthropic's acquisition of Bun in December 2024 signals the infrastructure side is lining up too.

00:04:06 And the practical engineering case Stagi makes is concrete: one language across agent loop, backend, and UI. Consistent typing with Zod throughout. One type checked once, used everywhere. NPM as an ecosystem for auth, payments, infra. Python still owns training, research, GPU serving.

00:04:27 But the application layer — where apps embed AI capabilities directly — is shifting. TypeScript doesn't own the brain of AI anymore. It's becoming the body that moves it. This connects back to review debt because both are loops running at different layers. The language agents default to will determine the layout of every code base built between now and the next model upgrade cycle.

00:04:55 And the review load those code bases create — the sprawl, the coupling, the architectural decisions made by people who only saw syntax — is what Gupta's metric catches. Whether TypeScript wins matters less than whether teams can build at that speed without creating a review burden their human capacity can't sustain.

00:05:18 Terence Tao used a coding agent to port twenty-four of his Java applets from 1999 to JavaScript over the weekend. He's a Fields Medalist mathematician at UCLA who's been interested in machine-assisted computation since Java 1.0 was current. The applets were visualizations for complex analysis, linear algebra, honeycombs, Besicovitch sets — stuff he originally wrote with Allen Knutson twenty-seven years ago.

00:05:48 His report is the kind of measured take you rarely see from someone in his position. He found one minor bug in the porting — a drag event that behaved oddly when dragged outside the main box — and the agent caught two bugs in his original code he hadn't noticed.

00:06:07 Net positive on accuracy for bounded work. He then asked the agent to build something new: a visualization tool for special relativity he'd tried to code in 1999 but abandoned because the complexity got too high. After a couple hours of what he calls vibe coding, he had it working.

00:06:27 He's also built one for his Gilbreath conjecture paper. The context matters here. These are secondary visual aids for mathematical arguments — not mission-critical components. The downside risk is acceptable. That's the bounded task envelope: well-defined output, low consequence of bugs, clear success criteria.

00:06:50 The local-model angle is that Tao's experience mirrors what Gupta's framework would flag as a low-review-debt scenario. The PR was small in scope, the author understood the goal before starting, and the agent worked within tight constraints. No architectural sprawl, no test theater, no evidence gap — because Tao reviewed each visualization himself.

00:07:15 This leaves one hangover for every bounded-task win: how does the unbounded case look? That's exactly where review debt compounds. An applet that renders geometry is a different problem than an agent writing the inventory system it just replaced.

00:07:33 Mario Zechner — libGDX creator, veteran game engine developer — put two thoughts on X this morning that cut through both the hype and the fear. First: control the types and interfaces. Let the rest fall into place. The latest models still love to introduce abstractions that fight your type system.

00:07:53 You need to beat them into submission rather than letting them stomp over things. Second: without constructing and understanding the overall system, you're fucked in the long run. But that doesn't mean you need to read all the code. We didn't do that pre-agents either on large projects.

00:08:13 False dichotomy. His point isn't against agents. It's about what you constrain them to work within. The type-and-interface boundary is a concrete engineering pattern — define your contracts, then let the agent fill in the implementation. The overall architecture stays with the human; the tactical decisions go to the model.

00:08:36 This maps directly onto Gupta's review debt framework. Most of Gupta's signal families measure whether someone understood what the code was doing before it shipped. Zechner's answer is essentially: you don't need to read every line, but you do need to own the boundaries.

00:08:55 The types are your boundary. The interfaces are your contract. Where Zechner and Gupta diverge is useful: one tells you what to do today — define tight interfaces — while the other tells you what to measure tomorrow. Teams that only do one of those will either write great specs nobody follows, or measure debt they can't act on.

00:09:18 Starbucks is building its own AI-assisted replacements for a Microsoft inventory system and an IBM maintenance platform. Bloomberg reported it this week. Starbucks spends roughly four hundred million dollars a year on software, and this is the first real signal I've seen of enterprise companies actually walking that back.

00:09:40 The market reacted by selling application-only vendors. IBM fell three percent premarket. ServiceNow dropped three point five percent. Salesforce slid four percent. Microsoft barely moved because it supplies the Azure infrastructure Starbucks will build on — they're still a customer there, just not for the application layer.

00:10:02 What's happening here is the buy-versus-build recalculation that everyone talks about until someone actually moves money. If your engineers already customize a vendor's product heavily to make it usable, and AI now lets them build a purpose-fit tool in a fraction of the time, the license fee stops looking like insurance and starts looking like rent you didn't know you were paying.

00:10:29 Starbucks isn't walking away from everything though. Earlier this year they pulled an AI-powered inventory counting system after it produced inaccurate counts and had to revert to manual tallies. They're learning before scaling, which makes the current move more credible than if they'd just announced it cold.

00:10:50 Sequence matters here: data consolidation, process redesign, and then AI. Teams that skip the middle step build expensive tools automating bad decisions faster. That's what Gupta's review debt catches at a micro level: when agents produce code before the team has redesigned the underlying process, you get sprawl, coupling, and trust that decays.

00:11:14 Four hundred million dollars is Starbucks' line item. Every Fortune 500 company now has to ask whether its internal tools are next.

00:11:23 These five items land on the same signal, just at different scales. The agents that default to TypeScript are producing code faster than review capacity can scale — Gupta's review debt is the interest rate on that speed. Tao's bounded task works because he grasped the overall system and stayed within it before starting — the opposite of uncontrolled sprawl.

00:11:44 Zechner's type-and-interface constraint is the boundary condition teams need if they're going to keep agents from creating architectural debt. And Starbucks proves the enterprise-side version of the same thing: when AI reaches the application layer, every rented system becomes potentially rebuildable.

00:12:02 The metric I'm carrying from this week isn't any single number. It's Gupta's point about two different kinds of speed — production speed versus trust speed — and the realization that teams treating them as the same number are borrowing against their own architecture without knowing it.