Hey everyone, picture a team that treats a new coding agent like a laptop refresh. Someone posts the install link. A champion records a five-minute demo. Leadership asks for adoption numbers by Friday. Two weeks later, a few people are shipping faster. Most of the org is still waiting for someone they trust to go first.
That pattern looks like a tooling problem on the surface. However, it behaves more like a social system wearing a software costume.
In July 2026, Murphy-Hill and coauthors at Microsoft published a field study of the company’s early-2026 rollout of Copilot CLI and Claude Code through about April 29, 2026 (arXiv:2607.01418). Around the same window, GitHub made GPT-6 Astra available for longer agentic coding work and gave admins model policy controls (GitHub Changelog, 4 Sep 2026). Cursor also shipped Projects for coordinator-led subagent work (Cursor, 10 Sep 2026). So the install buttons keep getting shinier. The hard part stays stubbornly human.
What the Microsoft study actually measured
The paper separates three questions managers usually mash into one slide. Who tries the tool? Who keeps using it? And does output move?
On output, adopters merged about 24% more pull requests than they would have otherwise. Using CausalImpact, the authors estimate a +24.0% lift in merged PRs per engineer per day (95% CI +14.5% to +33.7%, p < 0.001). That is a strong signal inside one large org. Still, it is not a universal promise for your company, your stack, or your review culture.
The authors stay careful about the proxy itself. They treat merged pull requests as a stand-in for output. They also say plainly that a merged PR is not the same as the value it delivers. So if you celebrate merge volume alone, you can score a successful agent rollout that ships more noise.
In weeks with five or more tool-use days, merged-PR lift was about +50.1% versus that engineer’s zero-use weeks. The lift did not clearly fade across roughly four months. Relative to each engineer’s own zero-use weeks, Copilot CLI weeks showed about +24.9% lift. Claude Code weeks showed about +11.4%. However, that Microsoft-specific split may be confounded by ownership.
The paper also relays a Fortune anecdote. One high user could have cost more than $1.4 million on the cheapest Claude Opus 4.6 pricing they cite ($5 per million tokens). Then read that as a cost-risk illustration, not a typical team bill.
Adoption travels through people you already trust
When skip-level peers hit 25% or higher usage, the odds of a first Copilot CLI try jumped by about +216%. Direct manager use associated with about +82% higher odds. Reviewer peers at 25% or higher associated with about +54% higher odds. First use spread through networks people already watch for norms. Instead, it did not spread through another all-hands deck about AI transformation.
Say you are an engineering manager staring at a stubborn plateau. Another mandatory enablement hour will probably move less than one respected skip-level engineer doing a live demo on a real ticket, with the awkward bits left in. People copy the path they see succeeding near them. They do not, for example, copy the path that only appears in a policy PDF.
That is also why “we bought seats” is a weak success metric. Seat counts mainly measure procurement activity. Meanwhile, visible peer use measures whether the social system has a pulse.
September 2026 product moves raise the ownership bar
Two product launches in early September make the social lesson sharper. The tools can now run longer without waiting for a human to type the next prompt.
GitHub’s GPT-6 Astra release targets long-horizon autonomous coding work on Pro+, Max, Business, and Enterprise plans at provider list pricing under usage-based billing. Enterprise and Business admins manage access through model policy. Under the default setup, new models turn on automatically unless admins disable the global default or that specific model. So if nobody owns that switch, default-on becomes an accidental culture change.
Cursor’s Projects feature puts a coordinator in charge of subagents. It runs in the cloud by default. It can also listen for Slack, schedules, or PR and CI signals. Cursor reports that new users merge 30% more PRs, while users who primarily use Projects merge six times as many. Treat those figures as vendor self-report, not independent science. Also, treat them as a playbook hint. Long-running agent work needs a named owner and a stop condition.
OWASP’s Top 10 for Agentic Applications (9 Dec 2025) covers the trust side through ASI09, Human-Agent Trust Exploitation. That risk is polished explanations that push humans to approve harmful actions. Agency without accountability fails in code review the same way it fails in any other approval loop. High peer adoption can raise confidence while lowering vigilance. So keep human merge gates on purpose.
A practical playbook for eng managers
You do not need another tool bake-off to start. Instead, you need habits that treat rollout as social design.
1. Make peer use visible before you schedule another install pitch
Pick champions across skip levels and review lanes. Then ask them to show one real workflow, including the failed attempt. Imagine a platform team whose first Copilot CLI demo is a refactor someone already planned to ship, not a toy repo.
2. Gate new models on purpose
When something like Astra lands under default enablement, decide who owns the model policy before the model owns your bill. Then write a short enablement rule covering teams, task classes, spend ceiling, and rollback. Default-on is a product convenience, not a governance strategy.
3. Name an owner and a stop condition for long-running agent projects
If a coordinator can chase a migration across dozens of PRs, someone human still owns done, pause, and abort. Picture a team spinning up a Project for a framework swap. The owner defines blast radius, review cadence, and the shutdown condition. Without that, autonomy becomes drift with better marketing. So name the stop condition first.
4. Measure outcomes beyond merged PRs
Keep the Microsoft lift in view beside the authors’ caveat. Track review rework. Track incident rate on agent-touched changes. Track escaped defects. Finally, check whether the work reduced cycle time for valuable outcomes. Merged PRs are a proxy rather than the product itself.
5. Keep human merge gates even when peer adoption is high
Social proof is rocket fuel for first try. However, it is also a trust exploit surface when an agent writes a confident rationale. Require the same accountability you would require from a strong junior hire, including clear diffs, test evidence, and a human who can explain why the change is safe.
AI coding agent rollout is culture work with a CLI attached
The Microsoft study suggests CLI coding agents are neither uniformly adopted nor a short-lived novelty inside that org. The September releases make longer autonomous loops cheaper to start and easier to leave unsupervised. Your job is not to win a model beauty contest. Your job is to design the social system around the tool: who goes first, who reviews, who owns the policy switch, what good means beyond merge count, and when a human still has to say no.
AI coding agents are tools inside software engineering, and they amplify the norms you already have. If those norms are clean, visible, and accountable, the tools help. But if those norms are install it and hope, you will get expensive hope.
If this article helped, drop a comment with the first workflow you want to try this week. Support the shenanigans buying me a coffee on Ko-fi, and follow Attune IT on YouTube or me on X.