Hello, my name
is Rohit Shukla

Infrastructure teams managed storage and virtualization platforms through separate, vendor-specific consoles - each with its own interface and approval chain. Provisioning one resource meant switching between three or four tools, a source of both delay and human error.

A multi-agent platform that unifies management behind a single natural-language interface: a coordinating agent delegates to platform-specific sub-agents, each translating intent into API calls.

• Central orchestrator + sub-agents - new platforms plug in without touching existing integrations
• Human-in-the-loop approval on every state-changing action
• RBAC enforced at the agent layer, not bolted on afterward
• A from-scratch MCP server where no existing integration offered enough control

Provisioning time dropped roughly 13× versus manual, multi-console operation - with every action still audited and approved.

The hardest part wasn't the orchestration - it was the approval layer. Agentic systems touching real infrastructure need guardrails built with the same rigor as the automation itself.

Model Context Protocol was new enough that no MCP server existed yet for a major enterprise storage platform's management API - every agentic workflow needed brittle, one-off integration code.

The first MCP server for the platform, mapping its REST API into MCP-compatible tools, and handling auth, error states, and rate limits so it was safe to expose to autonomous agents.

• Narrow, composable tools instead of one broad "do anything" tool
• Fail-safe defaults - destructive actions require explicit confirmation
• Schema-first design, treating the server as a contract

Became the integration layer for the orchestration platform above - and cut future integration work for that platform to near zero.

Designing for an agent consumer is different from designing for a human developer. Errors need to be reasoning-friendly, not just technically accurate.