An AI coding agent is a program that reads a folder, writes code, runs it and fixes its own errors, without a human approving every line. Claude Code, Codex and Gemini CLI are its three most visible faces in September 2026. Among technical teams, they have become the default tool. Among French small and mid-sized companies, they remain almost invisible. That gap is the real subject.
Tandem works with more than 40 companies on artificial intelligence applied to operations. This article reviews the coding agent ecosystem, with the numbers behind it. What the market gets wrong, what the agent does on its own, and what a management team can take from it without a development department. If you are looking for how to use one instead, our guide to Claude Code for non developers takes over.
What exactly is an AI coding agent?
A coding agent differs from an assistant through autonomy. The assistant suggests the next line or answers in a chat window. The agent takes an objective in plain language, draws up a plan, opens the files, runs the commands, reads the errors and starts again until the result holds. You describe the destination, not the route.
Three generations followed one another in five years. Code completion, which arrived with GitHub Copilot in 2021 and guesses the end of your line. Chat inside the editor, which explains and proposes. Then the agent, which appeared in late 2025 and executes. Today's leading tools all follow that third model.
- Claude Code, from Anthropic, available in the terminal, in the desktop app and on the web.
- Codex, from OpenAI, geared towards long tasks you launch and collect later.
- Gemini CLI and Antigravity, from Google, on the same command line principle.
- Cursor and GitHub Copilot, historically editors, which have added an agent mode.
- OpenCode, the open source option, for teams that want control over execution.
Andrej Karpathy, formerly of OpenAI, summed the shift up back in 2023 with a line I often reuse, "English is the new code". I devoted a full newsletter edition to Claude Code, non technical profiles included. The point that matters for a management team lies elsewhere. These ecosystems look more alike every month, and picking a tool commits you far less than it seems.
Where does coding agent adoption stand in 2026?
Among professional developers, the shift is done. The JetBrains Research survey of more than 15,000 respondents, run between May and July 2026, measures 90% weekly use of a coding agent and 68% daily use. Claude Code leads with 39% worldwide and 47% in the United States. Codex follows at 16%, against 3% in January 2026. GitHub Copilot slips to 21%, down from 29% a year earlier. The full figures are published by JetBrains Research.
Two lessons for a non technical management team. First, this market moves fast and no position is secure. Codex multiplied its share by five in six months. Second, the gap between first and fourth place stays modest given the noise around it. Choosing the tool is not the structural decision.
Why does the vendor report 97.9% and its clients 17.3%?
In June 2026, OpenAI published a paper on Codex usage that went round every corporate deck. The number quoted everywhere is 97.9% of its own staff using it, against around 40% in August 2025. The same document carries a second number, cited far less often. Among client organisations, usage stands at 17.3%. Among individual accounts, it drops to 0.7%. The Register reported both, noting that the data is self reported by the vendor and that it did not answer questions about possible internal incentives.
That reading changes the conversation. An AI lab is not a representative sample of the market, it is its extreme case. When a board hears that 98% of a company uses a coding agent, that is not the useful comparison. The useful one is 17.3%, and even that covers organisations already buying from the vendor.
What it changes for a company without a development team
The 2026 Baromètre France Num, run by Crédoc for the French Directorate General for Enterprise across 9,655 companies, gives the French picture. 40% of small and mid-sized firms use at least one AI solution, against 26% a year earlier and 13% in 2024. For mid-sized firms alone, the figure reaches 53%. The barometer is public and broken down by sector.
One figure from that survey deserves a pause. AI spending by these companies goes 17% to subscriptions and licences, 5% to training, 2% to consulting, and only 3% to custom development. In other words, 97% of them buy AI as off the shelf software. Almost none has anything built for itself.
That 3% line is exactly what a coding agent moves. An internal tool that never made it past the quote stage because it cost forty days of development becomes arguable again when it costs five. The dashboard nobody ever built, the in-house connector between two systems that do not talk, the form teams still fill in on a shared spreadsheet.
We applied that reasoning to ourselves. The site you are reading was rebuilt with Claude Design and Claude Code, and I tell the whole story in a dedicated article. When the need is a business application rather than a website, our custom internal tool offer describes the frame we set. And if the question is which project to start with, an AI audit exists precisely to settle it.
Does Claude Code replace n8n, Make and Zapier?
No, and that is the most expensive confusion around right now. A coding agent reproduces the logic of an automation scenario very quickly. It provides neither scheduled execution, nor logs, nor error recovery, nor a credential vault. Those four bricks are an automation platform's job.
I argued that position in a LinkedIn post in February 2026, as the wave of coding agent adoption was building. The field observation fitted in one sentence. Inside companies, a coding agent sometimes catches on faster than an internal chat assistant, because it produces something visible on day one.
The combination that works on our engagements uses both. The agent writes the business logic, the custom nodes and the interface teams handle. The platform hosts, runs and watches. If the subject is new to you, our complete guide to n8n lays the groundwork, and our n8n agency offer describes what we take on.
What a coding agent still cannot do alone
The demo is spectacular, the go-live less so. I filmed the build of a daily monitoring scenario from end to end, without writing a line of code. The agent found the bricks, assembled the architecture, spotted an invalid expression format and fixed it on its own. It also forgot the instruction for the writing model. Without proofreading, the email would have gone out empty.
Research backs what the field shows. The Google Cloud DORA report on the return on investment of AI-assisted development, published in May 2026, models a 39% first year return for a 500-person engineering organisation, with payback in around eight months. The same model raises the change failure rate from 5% to 6%. More code shipped faster saturates the controls already in place. The authors present their own estimate as highly uncertain, and a 500-engineer organisation bears no resemblance to an 80-person company.
AGENTS.md, the standard that limits lock-in
One question comes up in every scoping session. If we invest in a tool, what happens when we switch? The answer changed a lot in 2026. Vendors settled on a shared instruction file, AGENTS.md, an open format used by more than 60,000 open source projects and now stewarded by the Agentic AI Foundation under the Linux Foundation. Twenty-six agents recognise it.
Anthropic followed. The Claude Code documentation states that the tool reads an AGENTS.md directly from version 2.1.277, when no CLAUDE.md is present. In practice, the context work you do once, your conventions, your folder names, your writing rules, still holds if you change agent next year.
| When | The fact | What it changes for you |
|---|---|---|
| June 2026 | OpenAI publishes its Codex usage figures | The market benchmark is 17.3%, not 97.9% |
| May to July 2026 | JetBrains measures 90% weekly use among developers | Your suppliers already work this way |
| September 2026 | Claude Code reads AGENTS.md directly | Your context work survives a change of tool |
| September 2026 | France Num measures 40% AI use in small and mid-sized firms | The gap widens with the 3% who commission development |
Where to start without losing your nights
The first project decides the rest. A failure on a critical subject buries the topic for two years. Here is the order we apply with our clients, and it holds in five moves.
- Pick a visible, non critical deliverable. An internal dashboard, a monthly report, a documentation page. Nothing touching invoicing or payroll on the first attempt.
- Work in an isolated folder. The agent should only see what it needs. Permission modes and the risks that come with them are covered in our article on AI agent security.
- Write the context before you prompt. An AGENTS.md file at the root, with your conventions and brand constraints. Five minutes spent there save an hour of corrections.
- Proofread the result every time. The agent drops instructions without flagging it. Proofreading is not a precaution, it is a step in the process.
- Train two people, not twenty. A pair who genuinely master the tool produce more than a general awareness workshop. Our Claude Code training path describes what works for non technical profiles.
Two subjects run in parallel, not afterwards. Governance first, because article 4 of the European AI regulation has required AI literacy since February 2025, and we set out who decides what in a dedicated article. Budget next, because usage based billing holds surprises and the real cost per person deserves a calculation before you roll out.
Should you put an AI coding agent in your teams' hands?
Yes, on three conditions, and none of them concerns the choice of tool. An isolated scope, a deliverable that exposes nobody if it goes wrong, and two people trained seriously rather than a collective awareness session. The technology is mature. The open question is which project comes first, and that is settled internally, not in a tool comparison.
What would be a mistake is waiting. Not because you are behind, the numbers in this market say the opposite. But because the skill that counts is not knowing how to launch an agent. It is knowing which process deserves to be opened first, and that one is built by doing. If you would rather build it with support, a three-week AI audit gives you the map and the first deliverable. Our Claude training offer covers the other half of the subject.
This status report will be updated as the ecosystem ships. The next question we will document is the real cost at team scale, once the trial phase is over.
