Skip the marketing framing and read a vendor cookbook as a leading indicator of where production agent problems actually are. The recipes that show up first are the ones enough teams hit to justify writing a guide. On that reading, the Claude Cookbook is less a tutorial collection and more a field report on what breaks at scale.

Three entries stand out because they name problems I’ve been fighting in production, not features. Programmatic tool calling — letting the model emit code that calls tools instead of round-tripping every call through the message loop — is a direct answer to token-and-latency blowup when an agent orchestrates dozens of tools. Tool search with embeddings is the admission that once you have thousands of tools, you can’t stuff them all in context; tool discovery becomes a retrieval problem, and now your agent has a RAG layer whether you planned one or not. Automatic context compaction is the quiet acknowledgment that long-running agents drown in their own history.

The recipe I keep coming back to is the grade-and-revise loop: a writer drafts, a stateless grader fetches every cited URL and checks every quote against a rubric, and the feedback drives revision until it passes. That’s an eval harness wearing an agent’s clothes. The interesting shift is that verification is moving inside the loop instead of living in an offline eval suite you run after the fact.

None of this is novel if you operate agents daily — it’s the same reranking, budgeting, and observability work we do for RAG, relabeled for tool use. What’s useful is the signal: when a provider documents context compaction and tool retrieval as first-class recipes, those aren’t edge cases anymore. The HN thread argues about whether cookbooks lock you into one API shape. Fair. But the patterns outlive the API.

Which of these becomes table-stakes SDK behavior in a year, and which stays a recipe you’re expected to wire yourself?