A company buys coding-agent licences for the whole engineering team and expects use to spread by itself. Sometimes that is what happens. The question almost nobody asks is what exactly is spreading.
The thesis of this piece: if the use of coding agents travels between colleagues, each person's way of using them travels with it. Without a shared method, bad habits spread too. That is why it pays to start with one specific team that works in plain sight, with shared specifications and criteria, rather than with licences for everyone. From the study: first use spread mostly between colleagues. That the way of using the tools travels with it, and the rest, is onext's conclusion.
What Microsoft found in its rollout
Emerson Murphy-Hill, Jenna Butler and Alexandra Savelieva published a study in July 2026 on the rollout of two command-line agents, Claude Code and GitHub Copilot CLI, among tens of thousands of Microsoft engineers in the first half of 2026. It can be read on arXiv.
They find two things that matter here. First: "first use spread primarily through social networks". Second: adopters merged roughly 24 % more pull requests than they would have otherwise, and the lift held across the four-month window the study covers. From that they draw a recommendation: organisations should treat visible peer use as central to rollout strategy.
Limit It is a single study, published as an arXiv preprint with no peer review on record, about one company. It uses merged pull requests as the indicator of output, and the authors themselves warn that a merged pull request is not the same as the value it delivers. That is why the 24 % is not presented here as a productivity gain.
What was already known about how tools get discovered
The finding fits an earlier piece of work by the first author. In 2015, Murphy-Hill and colleagues studied how people who build software discover new tools: they interviewed 18 industry programmers and followed 76 users at their workplaces through a diary study (Computer Supported Cooperative Work, 24(5)). They concluded that learning a tool from a colleague works when it happens, but it happens rarely, roughly once every few months, and that it would improve if people talked about their tools more openly and concisely.
The two pieces say the same thing from different angles: word of mouth between colleagues is a real adoption channel, but it depends on use being seen and explained.
What travels with the tool
This is where what the studies say ends and what we think at onext begins. A colleague who shows how they use an agent is not only showing the command: they show how they ask for changes, how they check the result, how much they review and when they trust it. That is what others copy. If everyone works their own way, what circulates is a set of ways, with their wins and their shortcuts.
When use spreads by contagion, the tool is the least of it. What spreads is the method of whoever sets the example, good or not.
The difference between the two ways of starting, in a table:
| Question | Licences for everyone | One specific team, in plain sight |
|---|---|---|
| What spreads | Each person's way of working | An agreed, written way of working |
| What other teams see | Scattered results that are hard to compare | An example they can read and copy |
| How a bad habit gets fixed | Person by person, when it is found | In the team's method, for everyone at once |
| What can be measured | Use of the tool | Delivery and quality of one specific team |
onext's own reading of the Microsoft study (2026) and of DORA's capability on a clear AI stance (2025)
Where to start
There is no need to slow down anyone who already uses agents on their own. The point is that the example people see is the one you want copied:
- Pick a team. One with real work already planned and someone who wants to show how they work.
- Write down how it works. How a change is specified, which acceptance criteria close it and what context from the code the agent receives. In Spec-Driven Development (SDD), you write what the code has to satisfy first and build afterwards; it is explained in what SDD is and how it keeps code under control. Shared context is covered in Claude Code in teams: context, not control and in shared, curated instructions. A starting point is in the free SDD templates.
- Make it visible. Let the rest of engineering see the specifications, the reviews and the decisions, not just the result. It is the mechanism the study describes, but with a method behind it.
- Measure delivery and quality, not use. How long a change takes to reach production, how many fail afterwards and how much rework there is. What to measure is in the KPIs of an AI development team, and why to ask for a value thesis first is in the AI value thesis for the IT budget.
There is a precondition that DORA, Google Cloud's research programme on software delivery, lists as one of the seven capabilities in its AI model: the organisation's position on AI-assisted tools is clear and well communicated. That covers what is expected, what can be tried and which tools are allowed. DORA clarifies that it does not measure the content of the policy, only that it is understood (Storer and DeBellis, 2025). Without it, the contagion happens anyway, but with no norm to steer it.
And before handing out access, it helps to have the net that catches mistakes: before the agent, the safety net.
Before handing out licences
If you are rolling out coding agents and want to decide with someone from outside which team goes first, what it has to write down and how you will see whether it works, let's talk. The conversation starts with one specific team, not with the number of licences.
Frequently asked questions
How does the use of coding agents spread inside a team?
According to the study of the 2026 Microsoft rollout, first use spread mostly through social networks, and the authors recommend treating visible peer use as central to rollout strategy. It is a single study, about one company.
What does the Microsoft study measure, and what does it not?
It measures who adopts and keeps using two command-line agents (Claude Code and GitHub Copilot CLI) in early 2026, and whether their output goes up. It uses merged pull requests as the indicator of output and, as the authors point out, that is not the same as delivered value. It is an arXiv preprint with no peer review on record.
Does the 24 % mean teams are 24 % more productive?
No. The study says adopters merged roughly 24 % more pull requests than they would have otherwise, with the lift sustained over four months. A merged pull request is not the same as delivered value, and the authors say so themselves.
Why start with one team instead of giving licences to everyone?
That is onext's conclusion, not the study's. If use spreads between colleagues, the way of using it spreads with it. A team that works with shared specifications and criteria, and does so in plain sight, is the example worth copying. With licences for everyone, the example is whatever each person does.
What should the first team have written down?
How a change is specified, what the acceptance criteria are, what context from the code the agent gets and who decides where the risk is. It does not need a long manual: another team just has to be able to read it and copy it.
What is a clear AI stance according to DORA?
It is one of the seven capabilities in the DORA AI capabilities model: the organisation's position on AI-assisted tools is clear and well communicated (what is expected, what can be tried and which tools are allowed). DORA clarifies that it does not measure the content of the policy, only that it is clear and known.
Sources
- Emerson Murphy-Hill, Jenna Butler and Alexandra Savelieva, "Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI", arXiv:2607.01418, v1, 1 July 2026.
- Emerson Murphy-Hill, Da Young Lee, Gail C. Murphy and Joanna McGrenere, "How Do Users Discover New Tools in Software Development and Beyond?", Computer Supported Cooperative Work 24(5), 2015, pp. 389-422 (read on the publisher's record page).
- Kevin M. Storer and Derek DeBellis, "Introducing the DORA AI Capabilities Model", Google Cloud Blog, 23 September 2025.
- The table and the conclusion on what travels with the tool are onext's own work. Claude Code and GitHub Copilot CLI are named only because they are the two agents the study covers.
Bernat López is founder and CEO of onext, an AI boutique. He helps development and product teams work with AI with method —specification before building, a person who decides where there is risk and Spec-Driven Development— and applies to his own company what he proposes: onext runs on its own agentic system.
LinkedIn →