For Satya Nadella, producing code at low cost does not solve the difficulty of controlling it. The president and CEO of Microsoft also recommends that companies preserve their ability to change AI models.
Nadella wants controllable agents without technical dependence
Companies should view their AI agents as an internal risk, says Satya Nadella, interviewed by Gergely Orosz for The Pragmatic Engineer. Capacity is increasing, but reliability is not keeping up, he says, repeating an observation already expressed on 20VC last month. It calls for inviolable, auditable logging of the reasoning of these agents.
In his eyes, the now low cost of code generation shifts the difficulty towards its verification. Since the models used to examine this code can themselves suffer from misalignments, it favors cross-checking between distinct model families. He also wants to integrate the traces from production into a loop controlling the corrections, a device he refers to as “my RSI”.
This control also involves, according to him, through the separation between the software which carries the agent, its memory and the models requested, to avoid locking in with a supplier. Its criterion is to be able to substitute one model for any other while improving internal evaluations. A company incapable of doing so should, he believes, question its sovereignty.
Microsoft sees its interfaces evolve in step with the models
On the tools front, Nadella acknowledges that Microsoft was surprised by Claude Code, the Anthropic product that moved work from the editor to the terminal. Orosz questioned him about the lack of immediate response to his arrival in 2025, despite GitHub Copilot and VS Code. Nadella recalls that Copilot first offered to complete code; Cursor then offered a solid agent mode, which Microsoft eventually achieved.
Each output (GPT, Claude, Gemini, Mistral, etc.) decrypted the same evening, in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
The executive attributes the success of this command line interface to Anthropic’s model. About six months after this change, the integrated development environment, or IDE, had lost its centrality, according to him. He considers it necessary to quickly adapt interfaces to non-linear leaps in models, without being attached to what already exists. He nevertheless envisages that IDE will find a place again when decision models are integrated into software, comparing them to stored procedures for the intelligence era.
Nadella also says that the most advanced models from last year already run on a laptop, and those from last quarter on a desktop computer. He anticipates that routers like GitHub Copilot Fusion will distribute tasks between local and cloud execution. He presents this perspective as the beginning of hybrid intelligence and describes local models as “unmeasured intelligence.”
Microsoft plans to modify its teams before recruiting
Within Microsoft, the number of teams and their methods must evolve, announces Nadella, while the company employs 200,000 people. According to him, certain components have not yet integrated the transformation; the forward deployed engineer position is among those progressing the fastest. For young graduates, it sets a prerequisite: reorganize the teams and build their new “software factories”, then recruit. However, he considers it premature to extend the conclusions drawn from the current slowdown over time.
On the voluntary departure offers, Nadella did not respond when Orosz asked him if they preceded layoffs. He presents them as a means of granting more autonomy to employees, who are likely to choose functions that are economically relevant to them in the long term. He invites those who feel stuck to leave their team, highlights the existence of many internal positions and refuses to promise that nothing will change.
Microsoft also created PRAISE, a mentoring program for beginning engineers, with Carnegie Mellon University. For engineers, Nadella recommends reading less about AI and practicing it more. He says he wants to apply this rebalancing himself now, favoring action over speech.

