Microsoft CEO Satya Nadella stepped onto the Bengaluru stage not like a global tech leader carrying the weight of one of a $3.5 trillion company in the AI race, but like an engineer who simply couldn’t wait to show off a side project that he’d been hacking for hours.
Like some young developer out there exploring what wonders AI can do, Nadella has quietly been doing the same. His day doesn’t start with board meetings, but with AI assistant Copilot on his phone.
He wakes up, sets off agents, and then heads to work. He keeps coming back to it, only to see if the tasks been completed or what improvements the agents have made. “It’s just fun to be able to play with GitHub and just be constantly modifying,” Nadella said.
While the world was shutting off for Thanksgiving Day, Nadella started building what he calls an LLM Council, a multi-agent reasoning system. It has been inspired by Andrej Karpathy’s reasoning system that lets AI models debate inside an app.
Was fun to be at a dev event in Bengaluru and demo an app I built recently for deep research with multiple models and decision frameworks…think of it as "chain of debate"… Next stop, Copilot!! pic.twitter.com/oiinGN05BA
— Satya Nadella (@satyanadella) December 11, 2025
Creating one like a builder, he says, as a platform company, it’s always exciting to see the builders bring their creative and entrepreneurial energy. He even showed the app to Indian billionaire Gautam Adani, a day after meeting Prime Minister Narendra Modi.
Nadella shared that he works inside a Windows 365 environment using GitHub Codespaces, where his setup automatically generates several draft branches each morning. He reviews and deletes most of them by the end of the day, but usually keeps a piece or two that proves useful. Over time, those small daily experiments eventually grew into the decision-making framework he is now building.
On the model front, he said he relies heavily on Codex-Max because of its speed. Nadella said he wanted to go beyond building a simple deep-research tool, which pushed him toward the idea of the LLM Council. He even joked that all this work was really his attempt to get a job on the Copilot team.
What is LLM Council?
The LLM Council brings multiple models together. “We have all of these models available to you — GPT, Claude Opus, Gemini, Grok,” Nadella said. Users can pick one model to serve as the chair, with the rest acting as council members. The group then works like a selection committee, debating and weighing the query before producing an answer.
He compared the process to a chain of thought, but called it a ‘chain of debate’ instead.
For the demo, Nadella showed the council debated an all-time Indian Test cricket team, with different models offering contrasting judgments on openers, bowlers and captaincy. He explained that the system can spot problems such as era bias in cricket analysis, and that each model helps correct and refine the reasoning.
He then introduced DxO, a framework adapted from earlier healthcare research known as MAI-DxO. MAI-DxO stands for Microsoft AI Diagnostic Orchestrator, a model-agnostic multi-agent system originally built for medical diagnostics. “DxO is a thing that we implemented in healthcare first,” Nadella said. In this setup, Claude Opus leads the research, GPT-5.1 reviews for gaps and bias, and Gemini contributes domain and data analysis.
Nadella gave an example of a decision-making framework called Ensemble. In this setup, all the available AI models are used at once, but their identities are hidden. Instead of knowing whether a response came from GPT, Gemini, Claude, or Grok, the system anonymises them by assigning neutral labels like Alpha, Beta, Gamma, etc.
By removing model identities, the system prevents bias toward any specific model. All the responses are then combined, or synthesised, into a final single answer, so that the evaluation stays neutral.
Meta Cognition
Across all three frameworks, Nadella emphasised what he called a new form of user-level metacognition. “To me, this is the next generation of metacognition,” he said.
Although the agents produce research, critique and synthesis, he noted that “the metacognition is still us,” and these systems function as tools that strengthen human oversight.
His interface showed what he calls a “chain of debate,” allowing users to watch the models challenge each other and improve the answers before giving a final recommendation.
While Nadella used cricket as a demonstration, he said the same structures apply to complex decisions in supply chains, healthcare, finance and enterprise planning. “These types of chains of debates with multiple agents participating are going to be a lot of what we are all going to build,” he said.
Nadella also spoke about Agent365, Microsoft’s runtime layer for securing agentic systems. When he attempted to deploy his own agent into Microsoft’s tenant, the system blocked it.
He explained that Agent365 requires admin approval even for internally built agents. He described the runtime as essential for governance, observability and compliance in a world where enterprises deploy many autonomous agents.
He concluded by mentioning the momentum he is seeing across industries. Companies such as Cognizant, Persistent Systems, Swiggy and others are already building multi-agent orchestration frameworks. These examples, Nadella said, signal how rapidly agentic architectures are spreading through enterprise workflows.
The post Meet Satya Nadella, The Developer Hiding in Plain Sight appeared first on Analytics India Magazine.
