LLMs Drive Performative Productivity, Question Real Gains

Josh Collinsworth published a blog post on June 5, 2026 titled "LLMs and performative productivity," describing his firsthand experience using AI coding agents and questioning whether the activity they enable is truly productive. Per Collinsworth, agents sped up onboarding into unfamiliar codebases, a long-delayed Nuxt upgrade finished in about an hour, plus feature work, greenfield scaffolding, test creation, and numerous bug fixes. He reports that many outputs delivered little user-visible value, that he did not build enough context to defend some pull requests, and that several new projects or features were later abandoned.

Collinsworth contends LLM benefits are highly situational, helping most with boilerplate, greenfield work, and tasks outside the user's expertise, and helping experienced engineers less. He points to the limited body of objective studies on AI and developer output, arguing measured gains are smaller than they feel and can shrink, or go negative, once productivity is viewed holistically rather than by speed or volume alone.

Industry-pattern observations show a recurring tension between activity metrics and impact metrics. Organizations that gauge productivity by commits, pull-request velocity, or test counts may record short-term improvements that do not map to customer usage or system comprehension. When tools do the heavy lifting, higher output can coexist with weaker tacit-knowledge transfer, producing what the post calls a performative impression of productivity.

For teams adopting agents, useful signals include changes in code-review behavior, rates of reverted or flaky pull requests, usage telemetry for newly shipped features, and retention of institutional knowledge during onboarding. Dashboards that pair usage data with developer-output measures tend to reveal whether increased activity is translating into product value.