# Tool Access Does Not Aggregate Into Team Capability ## Claim Distributing a capability to individuals does not produce that capability at the team level unless a shared learning surface converts private practice into common practice. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds when a tool's effective use is tacit, individually tunable, and learned by trial rather than by documentation — the defining property of prompting-based AI tools. - Breaks when a tool's correct use is fully specifiable in advance (a linter, a formatter, a CI gate), in which case distribution genuinely is sufficient and no social layer is needed. - Doesn't claim individuals gain nothing from access — it claims the team-level capability organizations actually want (consistent quality, shared judgment) doesn't show up just because access exists. ## Evidence An analysis of 25,264 agent-generated pull requests across 2,361 GitHub repositories found 70% of projects had under one-in-five contributor participation in agentic workflows despite available tooling, and in 79% of agentic PRs the same developer both reviewed and modified the agent's output — no second human in the loop. Courtney Miller (George Washington University) names the mechanism: each engineer tunes their own prompting and correction habits, so "you can kind of end up in this situation where you just suddenly have very different processes, even though you started from the same place." Her diagnosis of the management failure: "If you internalize that, it's your responsibility to figure out how to speed up your productivity by 10x using these tools. That's a you problem" — organizations grant access and promise a multiplier, then leave the conversion to individuals rather than building shared Q&A channels, demos, or hack weeks that would let private practice become common practice. ## Links - [[AI & the Future of Knowledge Work MOC]] — why distributing a capability to individuals doesn't produce it at team level - [[3. Resources/Permanent Notes/Automating Both Ends Of A Review Collapses Its Value]] — because extends: the 79% self-review finding is a concrete instance of that note's broader claim that review only creates knowledge-transfer value when a genuine second human is present. - [[3. Resources/Permanent Notes/Goodharts Law Becomes Automatic Under AI Tool Use]] — because evidence-for: the article's closing metrics argument (PR count, then token consumption, treated as the same vanity-metric error) is a live instance of this pattern. - [[3. Resources/MOCs/Open Source Governance MOC]] — because applies-in: volunteer open-source projects rolling out new tooling hit the identical failure when access is announced without a teaching ritual. ## Tests / falsifiers - If teams given tool access alone reach the same participation rate as teams given tool access plus a shared learning surface, the claim is refuted. - If dispersion in agent-usage patterns within a team turns out not to predict output-quality variance, the "forked process" mechanism isn't actually the thing degrading team capability, and some other factor is. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/AI-coding agents kill team collaboration]] — reports the 79% self-review finding, the participation data, and Courtney Miller's/Sarah Wells's analysis this claim is drawn from.