AWS Labs just shipped a practical, agent-driven eval stack that can replace homegrown LLM bake-offs with reproducible, multi-judge results.
Use Opus 4.8 for decision‑grade long‑context work, and re‑test its coding chops before you lock in your next copilot contract.
Expect stricter MCP tool matching and rough edges around auth/metrics—fix your patterns, harden re-auth, and make telemetry null-safe.
Copilot isn’t only OpenAI anymore—Gemini 2.5 Pro is GA, so treat model selection as an engineering decision and test it on your stack.
Reassess RAG by default: CAG now works for many internal docs, and Qdrant is the pragmatic 2026 vector DB pick when you still need retrieval.