Attention Is the Budget
Part 7 of Code in the AI-Primary Era. Previous: One System, One Dynamics. This part is written for the people who sign org charts rather than commits; it stands alone if you skip the theory.
There is a syllogism making the rounds of planning meetings: our engineers have AI now; therefore productivity is up; therefore we can expect more, faster, from fewer people. It reads as obvious. It does not follow — and the gap between “reads as obvious” and “does not follow” is where the expensive mistakes of the next few years live.
The bottleneck AI exposes #
AI-augmented output depends on the whole process, and the process has a component AI does not accelerate: the human’s ability to understand what the AI did well enough to take responsibility for it. Review, integration, debugging code nobody on the team wrote, maintaining a mental model of a codebase that changes faster than anyone reads it — these were always part of engineering. They used to be invisible in the accounting, because other phases dominated the cost.
The structural change is easy to state. Acquiring ownership of code used to be fused with writing it. Typing produced understanding as a side effect — same person, same minutes, one line item. With generation, they split: producing the artifact is one project, and comprehending it is a separate project, displaced in time, performed against material its owner did not author. That second project was never measured, never optimized, never managed — there was no need. Now it is the long pole, and most organizations have no line item for it at all.
AI did not create this bottleneck. It exposed it — the way containerization exposed whose deploy process was held together by one person’s memory. The difference is that this bottleneck is made of human attention, which responds badly to the standard remedies.
You cannot pressure the AI #
Every engineering leader knows what deadline pressure does: the “soft” phases get squeezed — tests, refactoring, documentation. They get squeezed because their costs are deferred and their absence produces no same-day artifact. Under AI-primary development there is a new softest phase: the time spent acquiring ownership of AI output. It compresses for exactly the same reasons — no visible artifact, deferred consequences — and its compression is worse, because what ships without it is code that no one on the team ever understood, as opposed to code that at least its author did.
And notice where the pressure lands. It cannot land on the AI: generation speed and iteration count are properties of the model, indifferent to your quarter. When a deadline squeezes an AI-augmented team, the only compressible thing left is the human’s comprehension time. It will be squeezed by default, silently, unless leadership explicitly defends it — because no engineer, individually, can bill for “understanding what the agent wrote” while their dashboard shows the feature already green.
Attention economics #
The resource being spent in that comprehension phase is attention, and it is worth being precise about its properties, because they are unlike every other input in the budget. Attention does not scale with tooling. It does not parallelize. It recovers slowly once depleted, and degrades sharply under sustained load. It is consumed not only by review itself but by interruption, context switching, and — the novel entry — the per-output decision of whether to trust the AI’s last artifact, a small tax now levied on every interaction.
For decades the scarce input was engineering time, and we built the whole management apparatus — estimates, velocity, utilization — around time. The scarce input now is engineer attention, and almost nothing measures it. A platform, in this frame, has one job: conserve it. Surface what needs attention, suppress what does not, maximize the downstream value of every unit spent. A toolchain that burns attention casually — noisy diagnostics, opaque agent behavior, missing provenance, context that has to be re-established every morning — is not a productivity platform, however fast it generates code. That is the leadership-legible version of this series’ two platform goals: they are, in accounting terms, an attention-conservation program.
The open question, and the cost of pretending it’s answered #
Here is the uncomfortable part: the actual productivity of a human working with AI under adequate conditions — and the metrics that would describe it — is an open empirical question. Nobody has the number. The studies contradict each other; the vendor decks do not contradict each other, which tells you what they are. Decisions made before the number exists run on intuition, marketing, and headcount arithmetic — and the errors are not symmetric. They include: architecture shipped fast and owned by no one; liability formally accepted by engineers who were never given the means to comprehend what they signed; and burnout arriving precisely when the organization’s constraint became the attention of the people burning out. The money saved by cutting staff on the assumption that AI made the remainder more productive is routinely smaller than the cost of any one of those.
None of this is an argument against the productivity gains — they are real, and the organizations that capture them will outcompete the ones that don’t. It is an argument about what captures them. The winners will be the ones that treat engineer attention as the budget line, measure the comprehension phase instead of pretending generation is the whole process, and buy platforms on the strength of their feedback loops rather than their generation demos. Deadline math that ignores the owner’s ledger doesn’t make the debt disappear; it just books it where no one is looking.
This closes the descriptive arc of the series. What remains is prescriptive: the concrete requirements an AI-first platform must meet — machine-readable diagnostics, observable compilation, programmable extension, a structured data substrate — and an audit of how Logos’s architecture measures against them. Those parts are coming.