The autograder runs inside the student's machine

Sixty submissions, each a snapshot of a student container. The grading script copies each one and POSTs to the copy's agent URL: a headless run, one prompt, the rubric attached. Run the test suite. Read the code. Write feedback to /home/student/feedback.md. Score 0–10 per criterion.

The agent doesn't rebuild the student's setup. It's already inside it: their Python version, their requirements.txt, the sqlite file they forgot to mention. When the tests fail because the student hardcoded a path, the agent sees the path. When they pass, the agent reads the code anyway and flags the nested loop.

Tokens go through the Hoody AI Gateway, 5% over provider cost, drawn from the course's AI Balance. The copies are deleted afterward; the feedback file is copied back.

The grader stands in the student's environment instead of reconstructing it.