Agent builders wrestle with domain expertise

This is less a product announcement than a useful signal: multi-agent systems are running into boring production problems before they run into sci-fi autonomy.

• –Live data breaks the “just ask the LLM” pattern because pricing, availability, compliance, and API state need retrieval, normalization, and freshness checks
• –Domain agents make sense, but only if they expose structured contracts, provenance, confidence, and failure modes instead of free-form answers
• –The missing layer is an agent registry plus evaluation history: consumers need to know which specialist is trusted for which domain, against which tests
• –MCP-style tool interfaces help with access, but they do not solve quality, routing, observability, or reputation by themselves