Analysis of 15 production deployments reveals that LLMs select tools by reading descriptions, not schemas, leading to performance collapse with more than 10-15 tools for Haiku-class models. Five recurring patterns—Resource Gateway, Tool Orchestrator, Stateful Session Server, Proxy Aggregator, and Domain-Specific Adapter—address LLM-client constraints in naive designs. Tool selection accuracy for Sonnet 4 drops below 90% at 20 tools and both Haiku and Sonnet models are guessing at 30 tools.
Agentic system design is now a first-class architectural constraint, requiring specific patterns to scale beyond 10-15 tools and avoid performance collapse.