LESSON 8.6 · Engineering · 150 min
Post-training for agents
Post-training for agents. Learn tool-use trajectories, task synthesis, verifiable environments, and process/outcome rewards and complete: Design a replayable training and evaluatio
Learning objectives
- Explain what problem “Post-training for agents” solves without hiding behind terminology.
- Trace the variables and causal links across tool-use trajectories, task synthesis, verifiable environments.
- Complete “Design a replayable training and evaluation pipeline for a tool-using agent” and judge the result with evidence rather than intuition.
Core concepts
tool-use trajectories
tool-use trajectories is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment before moving to an optimized implementation.
task synthesis
task synthesis is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment.
verifiable environments
verifiable environments is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment.
and process/outcome rewards
and process/outcome rewards is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment.
Build and verify
Design a replayable training and evaluation pipeline for a tool-using agent
- Predict: write the expected output, trend, or failure before running code.
- Build: implement only the minimum components needed to answer the question.
- Verify: compare with a baseline or trusted implementation; save seeds, parameters, and raw outputs.
- Transfer: change one shape, dataset, scale, or workload condition and explain whether the conclusion still holds.