Do not train your own model yet, and respect the Bitter Lesson

This is one of the most important strategic lessons, and it ties back to the Bitter Lesson. The harness you build today will probably be made obsolete by the next frontier model. If you spend weeks fine-tuning a model or training a reinforcement learning policy for a specific action space, you risk locking yourself into a "local optimum": a setup that is good now but cannot ride the wave of improving general models.

The advice is to use context engineering as a flexible interface that adapts to rapidly improving models, rather than baking your current assumptions into trained weights. Boris Cherny, the creator of Claude Code, has said the Bitter Lesson influenced his decision to keep Claude Code unopinionated so it could adapt easily as models improved.

There is a practical test that comes from this. Run your agent's evaluations across models of different strengths. If a stronger model does not make your agent better, your harness may be holding the agent back ("hobbling" it). This is how you check whether your design is "future proof," in the language of the article. Hyung Won Chung's framing captures the mindset: add structure for the level of compute available today, then be willing to remove it later, because those shortcuts will eventually bottleneck further improvement.