Do social networks drive adoption of new coding tools?

The central claim is that Microsoft's early-2026 rollout of agentic command-line coding tools spread through social channels, that retention tracked what engineers do rather than who they are, and that adopters merged more output. The authors base this on developer-level telemetry covering tens of thousands of engineers over roughly four months. They report that "first use spread primarily through social networks," that "retention was associated more with engineers' coding activity than with demographics," and that adopters "merged roughly 24% more pull requests than they would have otherwise." The 24% is the authors' own counterfactual estimate from a within-person design. Their output measure is the merged pull request, which they concede is "not the same as the value it delivers."

The adoption study models first use as a discrete-time hazard on an engineer-week panel and defines retention as Copilot CLI activity on at least 5 of the 14 days after first use. The sample is Microsoft software engineers who could have tried Copilot CLI at rollout, after dropping engineers with Claude Code access. The strongest first-use predictor is whether peers, "especially the broader skip-level group," had already tried the tool. Retention runs on different terms: prior IDE Copilot use predicts trying the CLI but "actively predicts against sticking with" it. Career stage and tenure matter little, and the authors conclude that "what an engineer does (peer ties, prior tool use, PR cadence) explains who adopts and retains far better than who an engineer is."

The paper's main contrast is with prior method. The authors say a systematic review of 25 AI-tool adoption studies found none that used observational adoption data. Their enterprise roster supplies the eligible-but-not-adopting denominator that public-repository studies lack. Does generative AI shift knowledge workers away from communication? also reads workplace logs, but it measures a shift in task mix rather than tool uptake. The nearest observational-adoption neighbor, Where have workers actually delegated tasks to AI?, maps public GitHub agent skills onto occupational tasks. It works at occupation level, where this excerpt has one company's full population. The authors also set their non-fading PR lift against a Cursor study whose lift was gone by month three, attributing the gap to tool generation and to their within-person design.

The excerpt does not establish that the extra merged PRs mean better software. The authors say the field "still lacks agreed-upon measures" of quality and call that the pressing open question. The Claude Code comparison is only partly visible: Copilot CLI adopters show a larger PR lift, which the authors call surprising, and they offer two untested hypotheses, different task mixes and Microsoft's ownership of GitHub aligning the harness with internal work. The adoption study is cross-sectional, so peer influence cannot be separated from homophily. Sensitivity checks cover the retention threshold (3-of-14 and 7-of-14) and two sample-frame variants, but not the 28-day merge window. The implication, at the strength the evidence allows, is narrow: at one large software firm, visible peer use went with uptake, and adopters merged roughly 24% more pull requests over four months. It does not show that other organizations would see the same lift.