Overview
- Microsoft sets out a production architecture for agents—a plan→act→observe→refine loop bounded by budgets, stop conditions, and human‑in‑the‑loop gates—with interoperability work (A2A, MCP) as adoption reaches 35% with 44% planning, per MIT Sloan/BCG.
- Retrieval is engineered as a pipeline for grounding (hybrid search, ACL‑aware filtering, reranking, compact context), while action tools require least‑privilege credentials, input validation, idempotency, and post‑condition checks.
- Memory is treated as explicit state—working, session, long‑term—with policy‑aware retention and permission trimming across both stored state and retrieved enterprise content.
- Orchestration favors a manager‑plus‑specialists approach using strict JSON input/output contracts, robust failure handling, and a single authority for user‑facing responses to prevent conflicts.
- Real deployments underscore impact and controls: NeosAI saves up to 25 hours per legal case, JPMorgan’s COiN analyzes 12,000+ contracts with 150+ attributes, HR assistants cut 30–40% of live requests, Morgan Stanley expands advisor agents with rigorous evals, and a ServiceNow–Semantic Kernel prototype coordinates incident response.