AI agents in software development: what they speed up and where they get in the way
We use AI agents daily: Claude Code, Cursor and Codex. Our current product was built by four people with these tools, and the store page collected 1,071 views and 26 preorders between 25 August and 16 September 2026 with no promotion behind it.
The way that work actually runs is not the way it is usually demonstrated. Here is what holds up in practice.
Where the gain is obvious
Routine code. Components that follow an existing pattern, forms, tables, migrations, API wrappers. Work where the correct result is easy to recognise.
Exploring an unfamiliar codebase. When you inherit somebody else’s project, an agent answers in ten minutes the questions that used to take a day: where the entry point is, how authentication works, what happens when this form is submitted.
Tests, especially for code that already works. That is exactly the work people postpone.
First drafts of documents: estimates, specifications, task descriptions. The word that matters is draft, the one a human then cuts down.
Where we lost time
Architecture. An agent will happily propose a structure, and it will look plausible. The problem is that the cost of a wrong choice shows up months later while verification is impossible today. We make those decisions ourselves and write down the reason.
Integrations with undocumented systems. When the question is how somebody else’s accounting system behaves on real data, generation does not help. Access and a few days of testing do.
Tasks with implicit requirements. If the client has not decided what the result should be, the agent will produce a confident version that has to be redone. The speed up is an illusion: you get the wrong thing faster.
The rule we use
If the result cannot be verified in minutes, it cannot be accepted.
A practical consequence follows: the better a project’s checks, the more value agents add. Strict types, fast builds, tests on the critical paths, a linter. This used to be hygiene and is now a precondition, because these checks are what stop bad output before it reaches the product.
We hit the inverse case on our own website. The article template listed dozens of typography classes while the typography plugin was not installed at all. The code looked right, the build passed, and long articles rendered with default browser styles. No agent catches that kind of error. Checking the result does.
What changes is the process, not the code
More time goes into framing the task. «Build a contact form» produces something you will rewrite. A brief that names the fields, the error states and where the data goes produces something you can accept.
More time goes into review. That part is genuinely unpleasant: reading someone else’s code is always slower than writing your own. The saving only appears at volume, when there is a lot of routine work.
And a note on deadlines. Faster code generation does not shorten a calendar project proportionally, because it does not touch approvals, credentials and acceptance. We plan from useful developer output, roughly 120 to 130 hours a month, and agents move that number less than demos suggest.
Before inviting agents into a project
Collect the project rules in one file: stack, folder structure, naming, what must not be done. The same trick works with people: a list of prohibitions beats a description of the ideal.
Make sure the build and the tests run with one command. If verifying a result takes ten steps, nobody will verify it.
And agree internally on who owns accepted code. «An agent wrote it» is not a clause in any contract we know of.
FAQ
Do AI agents replace developers
They redistribute time. Less goes to routine code and to exploring unfamiliar projects, more goes to framing tasks, reviewing and verifying. Responsibility for the result stays with the person who accepts it.
Where is the gain largest
On repeatable work with a verifiable result: components built to an existing pattern, migrations, tests for code that already works, mechanical refactoring, and reading an unfamiliar codebase.
Where do agents create extra work
Wherever verification is slow: architectural decisions, integrations with undocumented systems, tasks with implicit requirements. Fixing a confident wrong answer takes longer than writing the thing yourself.
What does a project need for this to work
Fast builds, types, tests on the critical paths, a linter, and the project rules collected in one file. All of it was useful before, and with agents it becomes a precondition, because those checks are what reject bad output.