What to build instead

If you follow the principles, the simplest compliant design is a single-threaded linear agent: one agent, one continuous context, doing steps in order. Context stays continuous and nothing is lost. Yan says this simple approach "will get you very far."

The honest limitation is that for very large tasks, a single context window eventually overflows. His proposed answer is not to abandon the single thread, but to add a dedicated step that compresses the history of actions and conversation into key details, events, and decisions. He notes this is genuinely hard to get right, and that Cognition has gone as far as fine-tuning a smaller model specifically for this compression job. The benefit is an agent that stays effective over much longer contexts. He is upfront that you will still hit a limit eventually, and frames managing arbitrarily long context as a deep open problem.