Continually Improving Our Agent Harness
Cursor, April 30,2026
Cursor shares how they iterate on their AI agent harness through experimentation, measurement, and model-specific customization.
Cursor's engineering team approaches building their agent harness with vision-driven development, running experiments and using both public benchmarks and A/B tests to measure quality. They've evolved from static context engineering with guardrails to dynamic context that agents fetch on-demand, adapting to improving model capabilities.
The team measures agent quality through metrics like code keep rates and language model analysis of user satisfaction, while also tracking tool errors and degradations at scale. Their process includes automated monitoring systems and weekly Automations that identify regressions and surface issues, helping them drive unexpected errors down by an order of magnitude.
They also deeply customize the harness for each model's training specifics, such as using patch-based formats for OpenAI models and string replacement for Anthropic models.