Stanford Computer Science course releases strict Model Context Protocol guidelines for agentic code management

Defining the Agent's Role

Stanford's CS336 repository introduces a rigorous CLAUDE.md framework to ensure AI agents act as teaching assistants rather than just solution generators. The guidelines emphasize that for courses like CS336, which are implementation-heavy (Python/PyTorch), agents should focus on guidance, debugging, and code review.

Practical Constraints

Agents are explicitly restricted from:

• Completing TODO sections in student assignments.
• Refactoring large code portions into finished solutions.
• Implementing core assignment components like Triton kernels or training loops directly.

Instead, agents are instructed to ask clarifying questions about what the student has tried and point to lecture materials at cs336.stanford.edu.

Handling Errors

When a student faces an error, the agent is expected to explain messages from PyTorch, CUDA, or distributed training tools while encouraging the use of sanity checks and assertions. For instance, when a student struggles with a causal mask, the agent provides specific checks (e.g., verifying mask broadcast shape) rather than giving the fix. This configuration helps maintain the integrity of the educational process while still providing professional-level support through an active dialog.