What Agentic Coding Actually Changed in Our Shop: An Honest Accounting

The vendor content will not tell you where the tools fail. The conference talks are written by people selling nearby services. The developer discourse swings between worship and doom on a weekly cycle. So here is the thing we exist to supply at moments like this one. A working agency’s ledger, after a year of agentic coding tools in daily practice. Claude Code, Copilot’s agent mode, Cursor. Multi-file changes, test runs, step-by-step refinement, against real client work with real deadlines and real consequences. Run through the same honest-accounting format we used for distributed work in 2020.

The tools this year crossed the line the earlier assistants never did. From suggestion, the autocomplete era, to execution. The agent takes a task, works across files, and runs the tests. It iterates on failures and presents completed work for review. That crossing changes the economics of development genuinely, unevenly, and in ways that reshape the work rather than simply speeding it up. Both columns of the ledger, the reshaping nobody prices in advance, and the client-facing translation this openness exists to serve. In order, per the house format. The agents moved the bottleneck. Generation is cheap now. Standing behind the work is not.

Where the step-change is real

The gains column, with the categories named precisely, because precision is what the vendor content lacks. Migration scaffolding. The platform migrations we run constantly (content-model transforms, template conversions, API-version updates) contain huge tracts of well-specified, pattern-following work. The agents compress it dramatically. The Drupal 7-diaspora projects and the framework-version updates on our bench this year ran with agent-drafted scaffolding at a pace the 2023 estimates would not recognize.

Test coverage. The forever-deferred work, the test suites legacy sites never had, is now cheap enough to simply do. Agents generate credible coverage from existing behavior. Humans review and harden it. Sites we inherited untested are exiting the year tested, which makes every future change safer and cheaper.

Well-specified component builds. The design-system component with a clear spec. The hookup with documented endpoints. The agent builds a reviewable first pass in minutes, and the spec quality, not the coding, is now the schedule variable.

And legacy-code archaeology, the sleeper. The agent that reads an undocumented codebase and explains what it does. The inherited site’s mystery plugin. The 2014 custom module nobody understands. It has quietly become one of the highest-value uses in the shop. Comprehension work that consumed senior days now takes supervised hours. The common trait across all four: bounded, specifiable, verifiable work, which is the line the misfire column defines from the other side.

Where the tools misfire, expensively

• Vague requirements: the agents will confidently build the wrong thing from an under-specified ask. Fluency and fabrication from the same source, the structural finding from our first notes on these tools, now with runnable output. The cost is worse than nothing. Plausible wrong code that eats review capacity to reject. Spec discipline, always valuable, is now the difference between the tools compounding and thrashing.

• Architectural judgment: the decisions our engineering coverage has always taught. What to build versus buy. Where the boundaries go. What the maintenance decade will cost. The agents produce confident answers and no accountability, and every architectural misfire we caught this year was caught by exactly the senior judgment the tools cannot supply.

• Undocumented client context: the constraint that lives in nobody’s repo. The stakeholder politics, the brand nuance, the this-broke-once-in-2019 institutional memory. Which describes, candidly, most of what makes agency work agency work. And which is why the tools reshape our economics without replacing our function. The context is the product. The code was always partly commodity.

The reshaping nobody prices in advance

The finding that most needs writing down, because it is the one the adoption business cases keep omitting. Review capacity becomes the bottleneck. When generation speeds up and checking does not, the constraint moves. The shop’s real throughput this year has been governed not by how fast the agents produce, but by how fast senior engineers can review, harden, and stand behind what they produced. The inspection bench doubled in length while the workbenches sped up.

The compensating investments, priced honestly, because we made them. The QA discipline formalized. The review standards, the automated gates the CI runs before any human hour is spent, the audit services our own practice now advertises because the market’s need is visible. The spec practice upgraded. What the misfires taught us now filters what we take on.

And the junior pipeline rethought rather than abandoned. The entry-level work the agents absorbed was also the training work. The apprenticeship problem we flagged in 2020 compounds under agent tooling. Our answer (juniors reviewing agent output under supervision, learning the judgment tier directly) is an experiment we will report on honestly next year. The industry’s answer to it will shape who exists to do the senior work in 2035.

The client-facing translation

What buyers of development work should take from a working shop’s year, stated plainly. Development economics are genuinely shifting. Scoping assumptions, timelines, and team shapes are all in motion, and procurement built on 2023 day-rates for 2023 task lists is mispricing in both directions. The shifts are uneven in exactly the pattern of the ledger above. The specifiable compresses. The judgment tier does not.

And buyers deserve partners who can explain their AI practice specifically rather than wave at it. Which tools, on which work, reviewed how and by whom? What quality gates, and what is the client paying for at each tier? Those are the questions worth answering here, applied this time to our own development work.

The openness is the differentiator, and deliberately so. An industry mid-repricing rewards the practices that can account for themselves. The ones that cannot will hide the reshaping inside unchanged invoices until the market corrects them. We have documented that pattern at every tooling squeeze since the CMS-template era. Adaptation beat denial then, and disclosure beats both now. The ledger stays open, the QA bench stays long, and the juniors stay hired. The accounting stays honest. Ask your partners for theirs.