Cline's open coding harness expands beyond coding, usage up 20x

Cline has spent two years building an evaluation system based on real-world traffic, tuning its harness to work well on both large frontier models and smaller, cheaper ones, and it recently open-sourced its evaluations for open-weight agents. Huang framed the design challenge as modularity: giving the model the right tool and the right correction loop, and letting it detect and fix its own failures.

The usage figures sit alongside a broader shift Huang described. As agents and models improve, she said Cline is seeing penetration into more general knowledge work and into people who are less sophisticated in coding tooling, which is a different audience from the developers the harness originally served. Much of that inference runs on CoreWeave's serverless platform.

The stakes appear to hinge on whether that less technical audience actually shows up, and on how cheaply open-weight models can handle work that does not require a frontier model. Huang's argument is that control over both the model and the harness lets companies build custom tools for niche jobs that would not otherwise justify a big budget or a dedicated engineering team — a bet that may depend on how far that penetration trend continues.