The Inner Loop Your Coding Agent Quietly Broke

The productivity claim around coding agents is that they remove the typing bottleneck. The bottleneck the engineer actually hits in practice is different. The engineer can no longer hold the system in their head, because the agent is editing files faster than the engineer can read them, writing tests faster than the engineer can reason about coverage, and refactoring abstractions faster than the engineer can verify they still type-check at the design level rather than just the compiler level.

The tight inner loop — hypothesize, change, observe, refine — that defines competent engineering quietly collapses into a different loop. The engineer is now reviewing agent output rather than building intuition about the system. A METR randomized controlled trial from mid-2025 found experienced open-source developers were 19% slower on familiar codebases when using AI assistants, while reporting they felt 20% faster. The 39-point gap between perceived and actual productivity is not a measurement error. It is the sound of comprehension being silently traded for throughput.

The Inner Loop Is Not About Typing

Engineers who have shipped systems for a decade know the inner loop is not the keystroke cycle. It is the cognition cycle. You form a hypothesis about how a piece of state flows through three modules. You make a small change. You observe whether the system behaves the way your model predicted. If it does, your model strengthens. If it doesn't, you learn something you didn't know about the system, and your model updates.

The output of the inner loop is not the diff. It is the engineer who can now reason about the system at a level they couldn't an hour ago. The diff is a byproduct.

Coding agents accelerate the diff and decelerate the cognition. The change still happens. The hypothesis-test-update cycle does not, because the engineer is one rung removed: they are not forming the hypothesis, they are reviewing the agent's hypothesis. They are not observing the behavior, they are reading the agent's summary of the behavior. The signal that would normally update their mental model is filtered through an interpreter that has its own model and no obligation to keep theirs intact.

After a few months, the engineer notices that they can still ship features, but they can no longer answer architectural questions about modules they own. They have not stopped learning. They have stopped having anything to learn from, because the loop that taught them is gone.

Throughput Without Comprehension Is a Substitution, Not a Gain

The standard framing of coding agents is that they are a force multiplier. The honest framing is that they are a substitution. They trade a unit of comprehension for a unit of throughput. Whether the trade is good depends on what you need from the engineer next week.

The math is uncomfortable. If a team measures only velocity — PRs merged, line counts, story points — the substitution looks like a pure win. The 2025 Carnegie Mellon study reported developers spent 28% less time on boilerplate and 19% more time reviewing AI suggestions for complex logic. The shift is from authoring to evaluating, and evaluating is cognitively cheaper in the short term and cognitively poorer in the long term. You can evaluate something a thousand times without ever building the structural model that lets you originate something similar.

The team that doesn't price this substitution is decaying its engineering capability while celebrating the dashboard. Six months in, the codebase has grown but the team's grasp of it has shrunk. The most concrete symptom is when an engineer can no longer onboard the next hire, because they themselves cannot explain the modules they shipped. The agent could not transfer its model, because it never had one to transfer — and the engineer was never the source of the model in the first place.

What an Engineer Loses That No Dashboard Measures

The losses from comprehension decay are not visible until the moment they matter, which is usually an incident. Three patterns show up first:

Loading…