AI Architecture Updates: July 16, 2026

1. Domain-Specific Languages as a Reliability Layer for LLM Output

Martin Fowler / Unmesh Joshi. The article argues that constraining an LLM to a domain-specific language, rather than letting it emit unbounded general-purpose code, makes generation more reliable because the DSL strips away syntactic variation and ships with validators like parsers and compilers that let the model detect and repair its own errors. It proposes a two-phase pattern: use the LLM as a brainstorming partner to discover the right abstractions during design, then treat the settled DSL as a natural-language interface during implementation. The effect is to shift the source of truth away from fragile prompts and toward the DSL and its semantic model, keeping generated artifacts maintainable. Source

2. Agentic Resource Discovery Adds a Discovery Layer Beneath MCP

InfoQ. A new Agentic Resource Discovery specification targets a gap that protocols like Model Context Protocol leave open: MCP defines how an agent invokes a tool, but not how the agent finds that tool across organizational boundaries in the first place. Under the pattern, organizations publish machine-readable ai-catalog.json files describing their capabilities, registries aggregate those catalogs, and agents search by task intent rather than against static endpoint lists, with domain-based ownership verification to authenticate resources before connecting. The design is federated rather than a single global catalog and is positioned as a complementary layer alongside MCP and OpenAPI for enterprise trust and governance. Source