ABAmine Bouhlal

// Case study · 2026 — in delivery

A multi-agent service desk on Google's agent platform

NTT DATA — internal IT, enterprise scale

Solutions Architect — architecture owner

0

ITSM migrations — orchestrate above, never replace

A2A

peer agents: triage, identity, knowledge, and growing

4

models routed by task — Claude + Gemini via LiteLLM

WEB APPNext.jsTEAMS BOTEntra SSOORCHESTRATORADK · Agent RuntimeSESSIONS · MEMORYA2AIAM AGENTEntra ID · HITLKNOWLEDGERAG · citationsTICKETING PORTadapter / backendITSM BACKENDSsystems of recordPRESERVED
Web and Teams channels feed a triage orchestrator (ADK, on Agent Runtime), which delegates over A2A to specialised agents; a ticketing-port abstraction adapts to each ITSM backend, preserved as the system of record.

The problem

An enterprise service desk absorbing high volumes of well-understood requests — password resets, access, software — on top of ITSM platforms that encode years of process and compliance investment. Replacing them would be a multi-year migration with zero strategic upside; leaving the experience as-is keeps burning L1 capacity on routine work.

The system

Architected an orchestration layer above the ticketing backends, not a replacement: a triage orchestrator built with Google's Agent Development Kit (ADK), deployed to Agent Runtime on the Gemini Enterprise Agent Platform (the platform formerly known as Vertex AI, rebranded at Cloud Next '26). It classifies intent and delegates over the A2A protocol to specialised peer agents — identity & access (Entra ID), RAG-grounded knowledge with citations on every claim — with managed Sessions and Memory Bank for state, multi-model routing across Claude and Gemini via LiteLLM, and human-in-the-loop gates on sensitive actions. A ticketing-port abstraction keeps the agentic core decoupled from any specific ITSM backend, with an adapter per system.

The outcome

In delivery: eleven accepted architecture decision records, a multi-backend core that treats existing ITSM platforms as immutable systems of record, agent-eval gates in CI, and a design built to productise across business units — the reference architecture for agentic service operations.

Gemini Enterprise Agent PlatformGoogle ADKA2AClaudePythonNext.jsTerraform
Next case → Voice AI at logistics scaleDiscuss a similar system