Power users of AI coding tools are drawing on models from an average of 7.6 different AI labs in a single month, according to data from Kilo Code, an open-source agentic engineering platform. The findings, drawn from over 40 trillion tokens processed through the company's gateway, reveal that the most intense users don't rely on a single model or provider but instead switch between labs and models constantly, sometimes within the same working session. Brendan O'Leary, a founding developer relations engineer at Kilo Code, presented the analysis to show how early adopters actually use AI tools, in contrast to the industry narrative focused on capability benchmarks.

The data shows 92% of power users—defined as those making over 10,000 requests monthly—used models from multiple labs, compared to just 34% of light users who made fewer than 100 requests. Heavy users, making 1,000 to 10,000 requests per month, averaged 5.3 labs, while regular users at 100 to 1,000 requests averaged 2.9 labs. In April, 68% of developers who manually selected models used models from two or more AI labs in the same month, up from March. Forty-five percent switched models within the same hour during active work sessions. Among enterprise developers in managed environments, 71% ran models from two or more labs, a significant increase from the first quarter. The top 10 models by request volume came from six different AI labs, and no single model ever exceeded 37% of total requests on any given day.

According to O'Leary, the pattern represents "a learning curve, more than random experimentation," with developers discovering that different models deliver better results for specific tasks like frontend design, architecture planning, or database work. The report identifies three technical failure points that consistently break AI adoption: context construction, model routing, and feedback loops. Context construction—how files are indexed, which code is included, and whether the model understands project relationships—creates the most friction because problems here multiply as developers try to use more autonomous agents. Model routing matters because teams that assume a single best model will constrain their most sophisticated users and create incentives to work around the tooling. Feedback loops vary by adoption level: autocomplete is immediate, chat takes seconds to minutes, single agents take minutes, and orchestrated multi-agent systems can take hours, making trust harder to build at higher rungs.

O'Leary argues that adoption follows a four-rung trust ladder: autocomplete with ghost text, chat for code blocks and explanations, single agents that can read and write files, and finally orchestrated multi-agent systems that delegate across specialized capabilities. Each rung requires significantly more trust than the last, but the feedback loop gets longer and the cost of failure increases. Teams that jump directly to orchestrated agents without building trust at lower levels often fail, not because the agent performed poorly but because they designed no feedback loop to evaluate whether it was doing what they wanted. The report finds that developers who don't move past chat consume much higher token counts per request, manually pasting in context to patch retrieval failures. Trust is measurable through override rates for autocomplete, session depth and follow-up rates for chat, and completion rates for agents—whether developers let tasks run to completion or interrupt them. One highly visible agent failure in front of a skeptical senior engineer can create a narrative inside an organization that takes far longer to reverse than the time needed to accumulate positive experiences. Organizations should start rollouts with developers most patient with ambiguity and motivated to iterate, building social proof before expanding access. The multi-model pattern isn't slowing—it's accelerating with usage intensity, suggesting that infrastructure assuming a single provider will push sophisticated users to route around the tooling entirely.