2 min lesson
What a grader actually is
Respond to "Why is 'just use the test suite as the reward' insufficient for training a production coding agent, in one tight argument?" Name the reason and the detail behind it.
Step 1 of 3
What a grader actually isthe function that returns reward
A grader is the thing that turns an agent's trajectory into a scalar reward when no test can. It might be an LLM reading a diff against a rubric, a learned reward model trained on human preferences, an execution harness handing out partial credit or a blend. Whatever its form, it sits in the exact spot where the test used to be.
RL optimizes the agent to maximize whatever the grader returns. So the grader's ceiling is the agent's ceiling. A grader that can't tell a clean fix from a hacky one trains a policy that has no reason to prefer the clean fix. You are not scoring the model - you are defining, in code, what "good" means and the model will chase exactly that definition and nothing more.
This reframes the whole role. The interesting research is not a clever loss function. It is building a reward signal for a fuzzy human judgment, at a scale and reliability that survives millions of rollouts, on tasks where the ground truth is contested.
When an interviewer says "how would you train an agent to write better code," resist jumping to PPO. Lead with the reward: "What signal tells us the code got better? Tests cover correctness but not quality, intent or diff size and most tasks have no tests. So the first problem is building a grader I can trust." That answer shows you know where the difficulty actually lives.