Agentic coding adoption is social: testing a manager-and-team-artifacts hypothesis
Agentic coding adoption is social: testing a manager-and-team-artifacts hypothesis
Question. Can manager participation and shareable useful work help a team develop reliable agentic software practice rather than merely raising tool-use days? A good answer compares team-level interventions against individual rollout, follows actual artifacts into shared work, and counts time, errors, maintenance and delivered value along with uptake.
What we knew
Murphy-Hill, Butler and Savelieva, Adoption and Impact of Command-Line AI Coding Agents (July 1, 2026), model Microsoft engineers’ first use of Copilot CLI and their 5-of-14-day short-run retention separately. With over a quarter of engineers sharing a skip-level manager already using the tool, initial-use odds were +216% versus no exposed peers; direct-manager use corresponded to +82%. Corresponding retention odds were +66% and +22%. Peer influence and homophily cannot be disentangled; department-specific shocks and work selection also matter. Survey respondents described docs, issue analysis, prototypes and tools benefiting a team, but the survey cannot show whether those artifacts caused adoption or helped the business. The same paper’s 5+-use-day PR association is neither a use target nor evidence of net value.
Dru’s interpretation, September 29, 2026. He connected the peer-exposure pattern to a Netflix training comparison he heard: onboarding the whole team with its manager present was most helpful. He proposed manager-first learning, visible and praised useful team work (documentation, backlog triage, internal prototypes and small tools), then progressively more agent-handled workflow. This is his research hypothesis, not a claimed proven causal effect. The Netflix developer-platform session abstract and engineering-session abstract do not publish the whole-team-with-manager comparison; do not attribute a result to them. He and the group rejected making use frequency an adoption quota. ‘Internal’ does not automatically mean low risk: access, data exposure and reversibility define risk.
September 30 research brief and findings
Four cuts, one initial search apiece: (1) Is there team-randomized evidence comparing manager/team and individual coding-tool rollouts? (2) Which original operator accounts show manager participation and genuinely shared objects? (3) Can peer influence be distinguished from preexisting team selection, and is there counterevidence to compulsory shared use? (4) What outcomes distinguish adoption from net team benefit? Initial source list was the Microsoft paper and the two Netflix abstracts. Subsequent targeted checks examined the original trial registry, first-person engineering accounts and a field experiment outside software. No further query currently resolves the missing treatment results or finds the Netflix comparison published.
Direct experimental design, no result yet. The Siemens/HBS/LMU team-randomized trial registration plans about 500 teams each in control, individual-incentive and manager-plus-team-incentive arms. Primary endpoints are use, breadth and sustained use; secondary endpoints include proxy task-time telemetry, perceived productivity and knowledge sharing. It remained marked ‘In development’ with no results in the registry on September 30. It will not isolate manager participation from combined incentives, and does not preregister maintained-code quality or product value. Monitor results without prematurely claiming support.
Concrete operating mechanisms, not proof. Anthropic engineering director Fiona Fung says managers start as ICs and ship to learn the agentic workflow; pods decide how to run triage/planning while domain specialists keep review authority. Anthropic’s shareable PR walkthrough, incident-timeline and release-checklist artifacts are a product affordance, not adoption-effect evidence. lowtouch.ai’s founder instead describes account-specific deal folders, common marketing assets and pre-sales approval of agent-proposed CRM updates: concrete cross-functional objects and owners, but self-reported uptake with no matched comparison or error/benefit ledger. Distinct organizations, not pieces of a fictional feature chain.
Counterpoint. In an experimental study of 388 retail employees, mandatory paired joint AI use was associated with fewer and lower-quality documents than unstructured use; a different reframing treatment was associated with better upper-end quality. This is not coding or manager-led rollout, and its authors flag session-timing, attrition and grader confounds. It nonetheless counsels against substituting compulsory paired usage for visible useful outputs. The Microsoft observational signal can reflect shared project demands or selection rather than contagion.
Test design and remaining gap. Randomize teams (or stagger manager-attended onboarding with credible controls) to a specific artifact-and-feedback practice versus individual onboarding, holding access, learning time, manager incentives and task mix as constant as practicable. Predeclare an artifact’s consumer and check: e.g., triage suggestion accepted/corrected by PM, PR review packet actually used by reviewer, internal tool used and maintained by another teammate. Measure baseline and follow-up by task/risk tier: qualified uses, human preparation/review time, artifact reuse, independently rated utility, escaped bugs and corrections, inference spend, customer/CX feedback and later output at the same usage intensity. Record nonuse and adverse/abandoned attempts. Peer-network crossover requires measurement or cluster separation. Do not equate manager example-setting, team encouragement, compulsory co-use, visible artifacts and economic ROI; each is a distinct proposed mechanism.
Still missing. Published outcomes of the preregistered team trial; the original Netflix comparative claim; a causal agentic-coding study of useful shared artifacts with post-merge quality/customer outcomes; active human-time and error ledgers for the first-person cases. Next research only when one of those becomes available or Dru responds; meanwhile resume the single-feature trace.
Related: Microsoft productivity study; role accountability; economic evaluation; risk-tiered checks.