Super Agent Mania: Why the Agent Count in Your Enterprise Is About to Explode

There's a moment in every technology cycle where the consumer-grade version of the future outruns the enterprise-grade version of the present. We're living through one right now. 

That was the through-line of Rocky Lhotka's opening keynote at the most recent AI to ROI Leadership Series event. Rocky — VP of Strategy at Xebia, Microsoft Regional Director, MVP, and a builder who has spent a career inside the guts of enterprise systems — didn't open with a tools demo. He opened with a warning most leaders already feel in their bones but haven't said out loud: 

Your employees are not waiting for you. 

They have ChatGPT. They have Claude desktop. They have 18 flavors of Microsoft Copilot. They're pasting sensitive data into free tiers whose end-user agreements — the ones none of us read — quietly state that your data becomes training data. Whether you've sanctioned it or not, the shadow IT story of 1994 Microsoft Access and the early iPhone era is playing out again, faster and with much higher stakes. 

The Real Question Isn't Whether to Build 

Most executives are stuck in the same loop: the ground is moving so fast, if I build something today, won't I regret it in six months? 

Rocky's answer was blunt. Probably, yes. Build anyway. 

Because the alternative isn't a careful pause. The alternative is the Wild West your employees are already running through without you. And the organizations that wait another two or three years for the dust to settle won't just be behind on technology — they'll be behind on the experience of operating in a non-deterministic world. That experience cannot be bought, downloaded, or shortcut. 

The Anatomy of an Agent 

Before Rocky went into super agents, he stripped the hype off the word, “Super Agent,” itself. 

An agent is three parts: the LLM (frozen in time, does nothing on its own), the harness (the loop that sends, reads, decides, calls tools), and the directive (a markdown file telling the agent who it is). That's it. Every other term — copilot, assistant, bot, skill — is marketing on top of those three fundamentals. 

Every enterprise agent strategy starts with a harness decision. Most leaders don't realize they're making one.

The Super Agent Is a Portal Problem in New Clothes 

Now imagine your organization has 1,000 agents. Not 10. Not 100. A thousand — each acting like a highly specialized coworker, many taking autonomous action on your behalf. 

If you have 500 human employees and 1,000 agent employees, asking each person to mentally track which agent does what recreates the 1999 intranet problem at 100x the scale. 

The Super Agent is the answer. One front door. Behind it, a sales super agent routing to sales agents, an inventory super agent running the warehouse, an operations super agent watching the whole system and only escalating what truly matters.  

At Xebia, we call this architecture the Xebia Unified Intelligence Layer™. A Super Agent is the entry point. 

The Protocols Are Finally Real 

A year ago, agent integration was a mess. That has changed. 

MCP (Model Context Protocol), originally from Anthropic, is the escape hatch that lets agents reach your systems. Rocky's prediction: the next two to three years will see major enterprise investment writing MCP servers that expose core business systems to agents. 

A2A (Agent-to-Agent), lets peer agents discover each other and collaborate.  

2026 is the year serious enterprises can start building production agentic systems on MCP and A2A. 

What Nobody's Talking About Yet 

MCP Gateways for discoverability and security. Agent Identity in Entra ID — fine-grained permissions for non-human actors. Agent Registries so agents can find other agents. OpenTelemetry — because agents fail faster than humans can respond, which means agents must watch agents. Cost control — two cents a call is nothing, until you multiply by thousands of agents running continuously. 

What Leaders Should Take Away 

Shadow agents are already inside your building. You will have far more agents than you think. And experience cannot be purchased — the only way to develop instinct for non-deterministic systems is to build. 

The Agentic Enterprise isn't coming. It's here. It's just unevenly distributed. 

The only question left is whether yours is being architected on purpose, or by accident. 

Rocky Lhotka is VP of Strategy at Xebia. He delivered this keynote as part of the AI to ROI Leadership Series.