What is a browser agent harness?

A browser agent harness is essential for transforming a next-token predictor into a functional production agent capable of operating safely and efficiently in real-world environments. This concept, popularized by LangChain and operationalized by Claude Code, involves implementing a structured system around the model to provide security layers, caching, identity management, credential brokering, and skill memory. The harness acts as a context engineering tool that optimizes the model's efficiency by providing familiar tools, preventing context bloat, ensuring accuracy through a reasoning loop, and maintaining safety with guardrails and sandboxing. It compresses the complexity of running a browser agent into manageable layers, each addressing specific issues like prompt injection and caching, allowing the model to interact seamlessly with the web. The harness is designed to be small, editable, and opinionated, focusing on reliability and safety, which is crucial when deploying agents at scale for real users.