Module 3 · Reinforcement learning for agents · Lesson 6

Designing rewards

Turn code graders and a calibrated judge into a reward a policy can't cheaply satisfy, with weights priced from what each error costs, rubric and relative-ranking judges, and tests of the reward itself before any GPU time is spent.

80 min 2 videos 3 exercises 7-question quiz Answers

Setup for the exercises:Install Claude Code Python with uv Get an OpenRouter key Local models with Ollama OpenRouter or a local model?

By the end of this lesson you can

Compose a reward from a format gate, code-verifiable fields, a rubric judge and penalties, and derive each weight from the cost of the error it prices
Compute the decision threshold and the degenerate policies that a set of reward weights implies, before training
Choose between pointwise rubric or checklist judges and relative-ranking judges, and estimate a judge's cost per training run
Validate a reward before training with agreement on gold labels, a known-bad ranking test and a measurement of judge noise
Explain reward over-optimisation, and why a judge that is good enough for evals may not be good enough as a reward

In the last lesson's run, a triage record that missed the escalation on a $2,880 refund and promised the customer their money back scored 4 out of 5, the same as a record with the wrong workspace and more than a record that merely called the refund "billing". Nothing was wrong with the training. The reward said those errors were equally bad, and GRPO optimises exactly what the reward says. This lesson is about writing a reward that says what you mean: priced from what each mistake costs, with a judge only where code can't decide, and tested like any other piece of software before you spend GPU hours on it.

Designing rewards

Designing rewards

Designing rewards

Sign in to keep reading.

Sign in to keep reading.