NeuralOS
Agentes

Microsoft ships its Sales and Service agents to production on an MCP server with 90+ tools

Microsoft announced general availability of the Sales Agent and Service Agent across Microsoft 365 Copilot, Dynamics 365, Outlook and Teams. A purpose-built MCP server, with more than 90 tools, is what drives the Service Agent. The signal is clear: MCP has stopped being an experiment and is now production infrastructure.

EN
Equipo NeuralOS
Radar de IA
Jul 7, 20265 min read
In short

Microsoft put MCP at the core of a GA product, not in preview: its Service Agent runs on a custom MCP server with more than 90 tools. An agent's value lives in its catalog of actions, not in the model.

When an enterprise giant stops saying "preview" and starts saying "general availability," it's worth watching where it put its hands: it's no longer placing bets, it's billing. On July 7, 2026, Microsoft announced that its Sales Agent and Service Agent are hitting GA inside Microsoft 365 Copilot, Dynamics 365, Outlook and Teams. These aren't stage demos or closed betas: they're agents that work on companies' live CRM data, grounded in the real business information through a foundation of MCP tools and a layer Microsoft calls Work IQ. The detail the headline swallows is where they placed the critical piece.

## MCP stopped being the trendy experiment and became plumbing

The Service Agent reaching GA doesn't improvise: it's driven by an MCP server built specifically for service, with more than 90 new MCP tools. And they're not 90 generic filler utilities, but concrete business verbs that Microsoft details in its technical post from July 15: get a case, list them, update them, close them, enrich them with context, match against similar cases, reassign them, take the next one from the queue, resolve-and-move-to-next, or create a new case directly from an email. Each of those tools is an auditable action that the model invokes in a controlled way. MCP —the Model Context Protocol— was born relatively recently as an open standard for connecting models with tools and data; seeing it at the core of a Microsoft product in GA, and not in some lab corner, marks a phase shift.

## The value is no longer in the model, it's in the catalog of verbs

This is where the lesson that matters lives. Anyone can wire a powerful model to a chat. The hard part —and what separates a toy from a product a company lets touch its CRM— is the tool catalog: how many real actions the agent knows how to execute, with what permissions, against what live data, and with what guarantees it won't break anything. A model without tools is a brilliant consultant locked in a room with no phone: it gives gorgeous opinions, but it doesn't move a single case. Microsoft isn't showing off its model; it's showing off its 90+ verbs. That's the moat, and it's a moat you dig one tool at a time, not with a cleverer prompt.

## The moat is dug with integrations, not with demos

The pattern repeats across the whole 2026 agent industry: whoever has the widest and best-armored inventory of executable actions wins, because each new tool is one more task the agent closes end-to-end with no human intervention. And there's a security nuance Microsoft encapsulates well by building it as a dedicated MCP server: every sensitive action —closing a case, reassigning it, touching the CRM— passes through a controlled and audited surface, with per-role tool enablement, not a wild model access to the database. Catalog breadth and permission discipline aren't opposites: they're the two halves of the same moat.

## What it means for anyone building with AI

The moral for anyone building agents is uncomfortable in its simplicity: don't fall in love with the model, invest in the tool catalog and in how you protect it. It's the same ground NeuralOS has been digging for a while, with no smoke: a data-driven integrations engine, with a per-tenant encrypted Vault, OAuth, multi-account, and HMAC signing for the money paths in LatAm. The underlying vision matches Microsoft's —that the agent be worth what it knows how to do, not what it knows how to say— and MCP is the natural direction of that layer. When a giant puts MCP into production and bets its product on 90 verbs, it confirms something that sounds decidedly unglamorous: the boring work of wiring integrations well is, in fact, the work that decides the game.

Share
Ready to build?

Start building in
under 3 minutes

Join 4,200+ builders. No credit card. Build your first app with AI in minutes.