Signal
What to keep, what to skip. Commentary on the discourse around agent engineering: papers, tools, marketplaces, claims. Distinct from Field Notes, which report from inside the work.
Multi-agent is not an architecture decision. It is a workload decision.
Anthropic and Cognition were not disagreeing. One measured breadth-first research, the other warned about shared-state work. The variable is decomposability.
Long context did not kill RAG. It changed what retrieval is for.
Long context made lookup cheap. The harder work, selecting evidence, citing it, governing access to it, and keeping it current, did not move. Here is the routing rule and the literature that supports it.
The skills.md discourse is sharpening the wrong knife
The community is writing token-saving guides and building million-skill marketplaces. The actual constraint in Anthropic's Skills architecture is a 1% description budget that silently trims the labels of skills you invoke least.