Flagship · Microsoft · Shipped
Maverick
A multi-agent diagnostic platform that serves as the diagnostic interface for Azure VMware Solution incident response.
how measured
Steady-state monthly volume, across 8+ operational queues.
how measured
Time from incident intake to a triage decision.
how measured
Time from incident intake to the confirmed root cause. A separate metric from triage time.
The problem
Infrastructure incidents on Azure VMware Solution span many domains (networking, storage, compute, control plane), and diagnosing them means working through internal troubleshooting guides under time pressure. Maverick is the diagnostic interface for incident response: it takes an incident, investigates it, and hands a grounded diagnosis to a human.
The approach
A supervisor agent dispatches specialist sub-agents in parallel across infrastructure domains. Specialists pick tools dynamically and use retrieval over internal troubleshooting knowledge bases. Diagnostic capabilities are also exposed through a hosted MCP server, so other teams' agentic platforms can reuse them. I lead a team of 10 to 11 engineers building it.
It handles 500+ incidents every month across 8+ operational queues, and processes 200+ requests a day.
Key decision and tradeoff
The hard problems in agentic systems are determinism, reliability, and trust, not demos. So routing is deterministic and agents must execute the required diagnostic procedures before they are allowed to issue a verdict. Handoff to a human is confidence-gated, and nothing is remediated automatically.
Reliability
Long-running investigations are durable: orchestration is stateful and checkpointed, so a failure resumes instead of restarting. Specialists run in parallel, and context is managed so long workflows don't degrade. Evaluation gates sit between steps.
Status
In production at Microsoft.