By Guest: Randy Bias (LinkedIn)
Company: Mirantis
Show Name: To The Point
Topic: AI Infrastructure
The conversation around AI has been dominated by coding assistants and developer productivity tools. But according to Randy Bias, VP of Strategy & Technology at Mirantis, the most consequential shift happening right now isn’t happening in the IDE — it’s happening in operations.
📹 Going on record for 2026? We're recording the TFiR Prediction Series through mid-February. If you have a bold take on where AI Infrastructure, Cloud Native, or Enterprise IT is heading—we want to hear it. [Reserve your slot
Bias believes 2026 will be defined by a fundamental transition: AI agents that don’t just assist humans but autonomously manage infrastructure, respond to production events, and extend their own capabilities by writing and deploying code — all without waiting for a human to act.
Agentic Operations Is the Underfunded Frontier
The AI industry has poured enormous energy into AI-assisted software development. The operational side of the stack — networking, storage, bare metal, virtualization, cluster management — has received far less attention. Bias argues that the gap is about to close, and fast.
At Mirantis, the team is already working through the question in practice. With large-scale AI infrastructure deals on the horizon, Bias says the central challenge is clear: “How do we automate everything from top to bottom? How can the AI factory become something that can manage itself, for the most part?” The goal isn’t partial automation. It’s autonomous management of every layer of an AI factory — from bare metal to cluster orchestration.
MCP Becomes the Default Protocol
Bias is equally direct about where the agentic ecosystem is heading from a standards perspective. He sees the Model Context Protocol as the protocol — not one option among many, but the standard that wins by default.
“I’m convinced that this is the year that’s just a done deal,” he said. “I don’t think there’s really going to be any other protocols for most cases for agents. They’ll just keep extending the MCP protocol.”
That confidence is grounded in what Bias sees as MCP’s structural advantage: it gives general-purpose agents a consistent interface to domain-specific tools and capabilities through MCP servers, without requiring each use case to have its own custom agent built from scratch.
General-Purpose Agents Replace Custom Builds
This leads to one of Bias’s more provocative predictions: the era of custom-built agents is ending. Tools like Claude Code, Codex, Gemini, and Goose are already capable enough to tackle domain-specific problems when paired with the right MCP servers and skill sets.
“General-purpose agents can basically be put to domain-specific purposes by combining skills and tools in the form of MCP servers to tackle domain-specific problems,” Bias explained. The strategic advantage here is compounding — teams that bet on general-purpose agents get to ride the innovation wave from the underlying model providers rather than maintaining their own bespoke agent infrastructure. Bias estimates that 80 to 90 percent of agentic use cases can be handled this way.
Agents That Write and Deploy Their Own Code
The thread connecting all of Bias’s predictions leads to a single capability that he believes will define the next phase of agentic AI: agents that write code and push it to MCP servers to execute on their behalf.
A recent addition to the MCP protocol spec — the concept of tasks — makes this more concrete. Tasks enable long-running capabilities to be set on an MCP server and checked asynchronously, opening the door for agents to set their own execution logic and return for results without blocking.
Bias built a proof of concept to explore what this looks like in practice. He called it Nightcryer — a triage agent that is automatically dispatched when a problem occurs in a production Kubernetes cluster. The key detail is what triggers it: not a developer, not a human on call, but events in production itself.
“A general-purpose agent can basically write the code that has the rules and the actions that create the triggers for executing an automated action, like a triage event, and it can update that code on a regular basis,” Bias said. The implication is significant — MCP servers become dynamically extensible, shaped in real time by what agents decide they need.
For engineering teams building or managing AI infrastructure, Bias’s predictions sketch out a near-term future that demands new thinking about automation, protocols, and what human oversight actually looks like when the agents are writing the playbook themselves.
