Microsoft argues that improving what an agent retrieves, remembers and exposes to its model can reduce operating cost more sustainably than model selection alone. Its recommended stack combines focused knowledge retrieval, searchable tool catalogs and selective cross-session memory. Microsoft Azure Blog — 2 September 2026

• What changed: Microsoft is treating context engineering as an ongoing production-optimization process, supported by Foundry IQ, tool search and Foundry Agent Service memory.
• Why it matters: Sending an entire knowledge base, tool schema or conversation history on every request wastes tokens and can reduce answer quality.
• Practical takeaway: Measure retrieved-token count, tool-selection accuracy and answer quality separately. In RAG, return the smallest evidence set that still supports a cited answer.