In a November 16 interview on The Pragmatic Engineer podcast, Microsoft CEO Satya Nadella said advances in Anthropic’s models had made command-line AI coding tools a viable alternative to workflows centered on an IDE. He did not, however, expect IDEs to disappear permanently.
The discussion, reported on November 18, 2026, followed Microsoft’s October 26 announcement of a voluntary departure plan. Host Gergely Orosz asked whether the plan could precede further job cuts as AI improves productivity. Nadella did not make such a prediction. He said future opportunities would differ in team size, workflows and staffing, and described the voluntary plan as a way to give employees more choice. He also said Microsoft’s teams were not all changing at the same pace and that he could not promise everything would remain as it was.
From code generation to verification
Nadella said a jump in model capability can force AI products to change shape quickly. In his account, Cursor made effective use of agents within an IDE, while subsequent progress in Anthropic’s models made it practical to hand fuller programming tasks to a command-line tool. Orosz described developers moving away from IDEs over roughly six months; Nadella said IDEs had faded from view to some extent, not that they were obsolete.
Nadella called the emerging workflow a “software factory.” As generating code becomes cheaper, he argued, engineers need to put more emphasis on specifications, tests and verification: assigning bounded tasks to agents, checking the results and using failures to improve tests. He said this remains software engineering, but changes the level at which engineers interact with code.
Asked about slower hiring of junior engineers, Nadella cautioned that it was too early to treat a short-term slowdown as a lasting trend. Companies first need time to reorganize teams and establish new ways of working, he said.
Nadella also argued that businesses should keep their AI systems’ context, memory and orchestration separate from any single model so they can switch models and assess results against their own measures. For agents with the ability to take actions, he called for separating intelligence from authority and adding monitoring, audit logs and runtime safeguards. These were his recommendations for system design, not a claim that such protections are already in place everywhere.
