Anthropic published their thinking on context engineering — treating the context window as a scarce resource and designing retrieval, memory, and tool integrations around it. The most useful framing: agents should maintain lightweight identifiers (file paths, stored queries, web links) and dynamically load data at runtime rather than pre-processing everything upfront.

The piece on tool design is especially sharp. Bloated tool sets that cover too much functionality are one of the most common failure modes they see. If a human engineer can’t definitively say which tool should be used, the model can’t either.