IDC FutureTech Summit Germany 2025

Closed-door session on what changes when applications call each other through agents rather than fixed APIs. The interesting design question isn't which model — it's how identity, authorisation and observability work when the caller is non-deterministic. Walked through the agentic patterns ladder that frames the written piece: single agent first (still the right answer for most use cases), multi-agent only when context windows or genuine specialism demand it, and human-in-the-loop guardrails that loosen to human-on-the-loop as confidence builds. The protocol layer collapses the N×M integration problem — implement MCP once, connect anywhere — and with OpenAI, Google and Microsoft all adopting it inside twelve months, governance has moved to the Linux Foundation. That makes the boring infrastructure question — structured tool definitions, auditable tool calls — the place the operational risk now lives.