ENGINEERING LOG
How Microsoft's EvalAgentic project reframes AI cost management: context compression, prompt caching, model routing, and short-term memory. Applied to Justrite's meeting intelligence pipeline.
Read source ↗Post-generation validation: shape checks, evidence verification (span + quote substring), and format enforcement. "Not found" as a first-class output in enterprise RAG systems.
Read on Towards Data Science ↗Microsoft's open format for packaging domain expertise: SKILL.md + scripts + reference docs. Progressive disclosure in 4 stages. Three authoring modes: file-based, code-based, NuGet packages.
Read on Microsoft DevBlogs ↗Anthropic's latest recommendation: sub-agents can now spawn their own sub-agents (5 levels deep). CLAUDE.md as "compounding engineering." Parallel sessions as the real productivity multiplier.
Read on XDA ↗Five rules extracted from Anthropic's Fable 5 model: context density over volume, structured constraints over politeness, output verification loops, explicit fallback paths, and progressive refinement.