For a stretch the live argument was whether developers would use AI coding agents at all. It ended without anyone declaring it over. What replaced it was a different argument, about trust, and the timeline of the last eighteen months or so is the clearest way to see how one turned into the other.
About two years back: sceptical, and ownerless. A compilation of Stack Overflow, JetBrains and DORA results by Digital Applied puts developer trust in AI-generated code at 40% at the start of that window. In a separate enterprise data set from the same publisher, 11% of enterprises named a dedicated agent owner. Hold both figures in mind: most developers did not trust the output, and most enterprises had nobody named as responsible for it.
A year before the 2026 survey: Copilot leads. GitHub Copilot sat at 29% adoption among professional developers. It was the default because it was the one anyone had heard of.
January 2026: the shuffle begins. Claude Code was at 18%. Over the following months it more than doubled.
May to July 2026: the habit is universal. The JetBrains Developer Ecosystem Survey 2026, fielded with more than 15,000 developers, found nine in ten using an AI coding agent at work at least weekly and 68% using one daily. The shares by tool had reordered completely. They add to more than 100% because developers use several.
| Tool | Adoption share |
|---|---|
| Claude Code | 39% |
| GitHub Copilot | 21% |
| Codex | 16% |
| Cursor | 12% |
| JetBrains AI | 9% |
| OpenCode | 7% |
Source: JetBrains Developer Ecosystem Survey 2026. Claude Code went from 18% to 39% in the half year. Copilot fell from 29% to 21%. Codex rose roughly fivefold.
The same months: speed up, trust down. In controlled studies developers report finishing tasks 55-56% faster, per the Digital Applied compilation. In the same compilation only 29% trust AI-generated code to be accurate, down from the 40% two years earlier, and AI-coauthored pull requests carry more reported issues than human-only ones. Enterprise agent ownership went the other way, from 11% to 56% in two years. Companies noticed that someone had to decide what the agent may touch and who checks its work.
A figure we are withdrawing. An earlier version of this article subtracted two statistics to produce a 49-point "production gap": Gartner's finding that 80% of enterprise applications shipped or updated in early 2026 embed at least one AI agent, and S&P Global Market Intelligence's finding that 31% of enterprises have an agent running in production. We are retracting the subtraction. The two figures come from different firms with different definitions, one counting applications and the other counting enterprises, and "embeds an agent" is not the opposite of "runs an agent in production". Each number is fine alone. Together they measure nothing.
What the trajectory implies
Read the timeline as one line. Output got cheaper, confidence fell, the leading tool changed in a year, and ownership caught up last. Writing code was never the slow part of delivering software that matters. Reviewing it, testing it and being accountable for it were, and the agents made the cheap part cheaper. A team that celebrates faster output without growing review capacity has built a queue.
The tool churn carries its own lesson. If the leader can change inside a year, a plan that treats one vendor's agent as a fixture will not survive it. Build the review process, the access limits and the test gates around the work, and treat the agent as a part you can swap.
If you commission custom software, your suppliers are already in this timeline. For a company running Odoo or ERPNext, much of the custom work is code: modules, reports, integrations. Ask any implementer, ours included, who reviews agent-written code and how ("a developer looks at it" is an answer, "the agent checks itself" is not), where agents stop (money, compliance and customer data deserve line-by-line human review), what is tested on staging before your live system, and what access the tooling holds. An agent with a production credential is an injection risk, as our piece on prompt injection explains. For what stalls pilots, see agentic AI implementation mistakes.
ThinqHub implements Odoo and ERPNext, and our approach puts delivery discipline ahead of whichever tool is fashionable this quarter. If you are scoping custom work and want the review and testing model agreed before the first line is written, talk to us.
