Agent Post-Training Systems
Build an agent-training plan from trustworthy traces, teacher feedback and verified outcomes. Audit token masks, price distillation, preserve reward meaning and bound asynchronous rollout queues before approving a run.
An agent-training system turns model actions, tool observations and task outcomes into parameter updates. Each boundary can change what the model learns. A tool reply mislabeled as an assistant target changes the loss; a failed verifier labeled as an unsuccessful task changes the reward; an unbounded rollout queue changes which policy supplied the training data.
This course follows a constructed repository-repair agent through those boundaries. The first module produces a training proposal, a trace audit and a teacher-scoring budget. The second turns those artifacts into reward rules, a capacity plan and a launch review. Calculations and CPU exercises are teaching examples, with their assumptions exposed. They do not claim a trained model or measured GPU speedup.
The opening lesson is free. The five deeper lessons, including the launch-review project, require Premium. Each includes a worked example, an original mechanism diagram and self-checks. You can complete the reasoning exercises without a GPU.
For a full implementation-specific environment walkthrough, continue with the AgentGym reinforcement-learning guide. For a small executed model-training experiment, use the tiny language-model post-training project.
When you finish, you can
- Choose a training method from the failure and the supervision available
- Audit a multi-turn trace after tokenization, masking and truncation
- Budget teacher scoring separately from student generation
- Separate task failure, exhausted task budget and infrastructure failure
- Size rollout concurrency from admitted learning work and policy age
- Defend a launch decision with a reproducible evaluation and rollback plan
BEFORE YOU START · Comfort with Python, next-token prediction and a serving memory budget. Foundations of AI Infrastructure and the training-state lessons in Training at Scale provide the background. No reinforcement-learning research experience is assumed.
Syllabus
Choose and audit the training signal
Start from a reproducible agent failure, inspect the targets created from its trace, and budget teacher feedback on student-generated states.
Keep the rollout system honest
Preserve task outcomes through environment failures, control queues and policy age, then defend a launch using the complete evaluation population.