90% of developers now use AI coding agents. Here is what buyers should demand.

Six months ago we wrote that AI had broken the old build-vs-buy math. The cost of well-scoped custom software had collapsed. What we did not yet have was a clean picture of how developers were actually working.

JetBrains just published that picture. Their Developer Ecosystem Survey 2026 — more than 15,000 professional developers, May through July — found 90% using AI coding agents at work at least weekly, and 68% using them daily. This is not autocomplete. Agents plan, edit across a repo, run commands, and open pull requests. Claude Code is now the most-used coding tool at work (about 39% worldwide, 47% in the US). Codex grew roughly 5x in half a year. GitHub Copilot, the tool that introduced the category, is no longer the leader.

If you buy custom software, this is the most important delivery-market fact of 2026. Everyone serious is using these tools. The question is no longer whether your vendor is "AI-powered." It is whether they are using agents like a senior team — or like a slot machine.

What changed since the autocomplete era

Copilot-era tools suggested the next line. Agent-era tools take a ticket and come back with a diff.

That is why the build-vs-buy spreadsheet moved. Implementation time for a well-scoped workflow really did compress. It is also why a new failure mode showed up in the same breath: software that looks finished in an afternoon and is not owned by anyone who could explain it on Monday.

Google and Kaggle's five-day "vibe coding" intensive — 353,000 registered learners, recapped this month — named the graduation problem out loud: take prototypes from vibe to live. The phrase is doing honest work. Natural-language programming is a real skill. Production is a different one.

Vibe coding vs. production (the distinction buyers should use)

If a demo arrives in 48 hours, be impressed — and ask what happens in week three. Speed is the easy part now. Internal tools still die from neglect, not from slow typing.

What you should demand in the statement of work

You do not need to pick Claude Code vs. Codex vs. Cursor. That is an engineering preference, and it moves fast: GitHub Copilot led this market a year ago at 29%, and Cursor slid from 18% in January to 12% by summer.

Whatever tops that chart the week you sign will probably not top it the week you launch. You need process that survives the tool.

1. A named reviewer, not a named model.
Who reads the diff before it hits your repo? If the answer is "the agent," you do not have a vendor. You have a generator.

2. Tests and a demo of the unhappy path.
Happy-path screenshots are cheap. Ask to see auth failure, an empty state, and what happens when the third-party API is down.

3. You own the repo on day one.
Source, CI, and documentation in your account. Agent transcripts are not a handover. If the relationship ends, you should still be able to hire the next team.

4. Secrets stay out of the chat.
Production credentials do not belong in a prompt window. Keys are short-lived, environments are separate, and the agent does not get production write access "to debug faster." (If you are also putting business agents into production, the same instinct applies — we wrote the operator checklist earlier this month.)

5. A maintenance number that is not zero.
AI did not delete the 10–20% of build cost per year that keeps software alive. It deleted some of the typing. Budget the thinking.

6. An honest scope, still.
Agents do not fix a fuzzy workflow. They will build the wrong thing faster. The one-sentence test still applies: when X happens, the system does Y, and a human steps in when Z.

Ticket → Agent draft → Human review → Tests / security pass
↓                         ↓
Reject / tighten            Merge to your repo
↓
Demo the unhappy path, not just the screenshot

How to read a quote in 2026

A quote that is much cheaper than last year's should come with a sentence about how the work is reviewed. A quote that is the same as 2023, with "we use AI" in the deck, is someone capturing the margin instead of passing the speed on.

Neither extreme is automatically honest. Ask:

• How many hours of senior review per week are in this number?
• What is explicitly not in v1?
• Who is on the hook when the agent introduces a dependency you did not ask for?

OpenAI cutting frontier API prices this month (GPT-5.6 Sol, a temporary 20%+ drop) will make agent runs cheaper still. Token price is not quality. Cheap generation with no review is how you buy a rewrite.

What this means if you are the buyer

Use the speed. Four-to-eight-week first releases for a narrow workflow are a reasonable default now, not a heroic claim. Use the skepticism. The market is full of people who can produce a working-looking app and cannot operate it.

The firms worth hiring look slightly boring on a sales call: they talk about scope, tests, ownership, and what they will not automate yet. They use agents all day. They just do not confuse the agent's velocity with the team's judgment.