Resource · Technical manual
Agentic systems on critical data.
How to design, build and run AI agent systems when the data is regulated, confidential or patient-related. Current as of August 2026 (the MCP 2026-07-28 spec, A2A v1.0, ADK 2.0 and the OWASP agentic Top 10), grounded in a pharma company as the running example.
30 pages · 5 figures · August 2026 edition · PDF in Spanish
What's inside
From protocol to worked example
Protocols: MCP and A2A
What each one standardises, what the 2026-07-28 spec changes and where to draw the security boundaries.
Orchestration frameworks
ADK 2.0, CAMEL and LangGraph, with criteria to choose without locking yourself in.
Reference architecture
Layers, enforcement gateway, HITL gates, quarantine for untrusted content and per-task budgets.
Identity and least privilege
Workload identity, short-lived tokens, per-tool permissions and egress control.
Execution environments
Ephemeral sandboxes, gVisor and microVMs, typed tools instead of a shell, and where isolation stops protecting you.
Threat model
The OWASP agentic Top 10 (ASI01 to ASI10) and MITRE ATLAS, with concrete controls per risk.
Secure RAG
Query-time ACLs, minimisation, data provenance and sovereignty.
Compliance
The AI Act with the post-Omnibus timeline, GDPR, NIS2, ISO 42001 and the GxP layer for pharma.
Worked example with code
A complete pharmacovigilance triage: schema, quarantine, MCP server, graph with HITL and the poisoned-PDF test.
