The experiment involved a coding agent tasked with developing a text search service in Rust. The researcher added context in stages, tracking how each addition altered the project implementation plan rather than evaluating the final code quality.

A five-point scale was used to measure the impact of context. A score of 1 indicates minor cosmetic changes, while a score of 5 represents significant changes, including the emergence of alternative implementation paths that require human decision-making.

The process was divided into two phases: planning (steps 1–14) and execution (steps 15–24). High-impact information, such as corporate development rules and existing internal repositories, forced the agent to restructure the plan and adopt specific internal frameworks.

Low-impact information, scoring a 1, typically included clarifying questions about already established decisions or repetitive audit requests. These inputs did not change the core plan but merely added explanatory documentation or confirmed the absence of new issues.

The study highlights that context effectiveness depends on the role of the information. Data that cannot be inferred from a technical specification, such as internal team agreements or specific repository structures, has the most significant influence on the agent's reasoning.