MentorScript is the artifact Structured Agentic Software Engineering (SASE) (SASE) proposes for codifying team norms and best practices for agents, introduced in Agentic Software Engineering: Foundational Pillars and a Research Roadmap as the product of AI Teammate Mentorship Engineering (ATME) (ATME). The paper frames this as transforming mentorship from an implicit, ephemeral activity — comments in a code review — into an explicit, evolving, codified discipline it calls “mentorship-as-code.”
Usage
A MentorScript can hold rules ranging from granular checks (e.g. “all new functions must have deterministic tests”) to high-level principles, and the paper states such rules should be subject to their own quality gates — linting, unit testing, and conflict detection — to keep them atomic and deterministic. Guidance can be explicit and durable (a direct, generalizable correction a human coach captures so the agent does not repeat a mistake) or inferred (an agent proposing a new general rule from a specific contextual correction, for the coach to approve). The paper states every action an agent takes should be traceable back to the MentorScript rules that were considered, using prompt-interpretation and reasoning-observability techniques, to enable rapid root-cause analysis when an agent’s behavior deviates from expectations. It cites project-level configuration files such as CLAUDE.md, .clinerules, and AGENT.md — see AGENTS.md — as an early, grassroots example of MentorScript-like artifacts, and notes the community currently has no consensus on what such files should contain or at what level of detail.
Related Terms
AI Teammate Mentorship Engineering (ATME), AGENTS.md, BriefingScript