Ask whether you should train before asking how.

A training algorithm is not the beginning: high-quality data, an interactive environment and a credible reward determine what the model learns.

Fix context and tools first; train only behavior that truly needs to become a parameter.

01 / CHOICEChoose the problem to solve first

Post-training is easily mistaken for a universal ability upgrade. Reverse the order: stabilize prompts, context, tools and evaluation, then ask whether the issue really requires changing parameters. If knowledge retrieval, a clearer instruction or a deterministic validation fixes the behavior, training adds cost and uncertainty without solving the actual constraint.

02 / STAGESPre-training, SFT and RL solve different jobs

Pre-training builds general language and world knowledge. Supervised fine-tuning (SFT) uses high-quality examples to shape behavior and output format. Reinforcement learning (RL) optimizes a policy from reward in an environment. A stable SFT baseline normally comes before RL; confusing these roles creates data projects with no measurable target.

03 / SFT OR RLUse SFT for demonstrations and RL for reliable rewards

When experts can demonstrate correct behavior and the target format is clear, SFT is direct. RL becomes valuable when many solutions exist, outcomes can be automatically checked and the system must improve beyond examples. An unreliable reward teaches metric gaming; an unrealistic environment yields a policy that cannot transfer. In multi-turn work, final failure may come from an early tool choice: outcome reward is objective but sparse, while process reward is dense but can narrow exploration.

04 / READINESSPrepare training before tuning parameters

Before training, confirm baseline evaluation, data provenance, privacy permission, resettable environment, reward anti-gaming, rollback plan and cost budget. Without these, parameter tuning only amplifies uncertainty. First exercise: take one agent failure and write a prompt repair, tool repair, SFT and RL option; explain why one comes first. The answer should follow evidence and reversibility, not fashion.

Sources

bojieli/ai-agent-book.
Chapter 7 — Post-training.
This is an applied reorganization, not a replacement for the source book; use the source text and code for exact claims.

Next: how an agent can evolve continuously.

Continue to continuous learning