Dr. Elisha Rosensweig is a computer science PhD, a researcher at Dicta - the Israel Center for Text Analysis, and the creator of the podcast "Elisha and the Angles." He has worked in high-tech at the intersection of development and research for over a decade, specializing in training R&D teams to work with AI tools in ways that empower rather than confuse them.
The Problem with "Try the Tool and See"
Most teams adopting AI coding agents do so through trial and error. Someone recommends Cursor, another tries Claude Code, and after a week everyone is working differently with no shared language. Rosensweig argues this is not a tooling problem, it is a methodology problem.
An AI coding agent is not just another editor plugin. It is a cognitive partner that needs to be managed in a structured way, with clear instructions, consistent quality control, and a review process built into daily work. Without methodology, the AI produces code that looks correct and can break things in ways no one anticipated.
What a Developer Using AI Coding Agents Correctly Actually Does
Rosensweig describes several principles he teaches in his workshops. The first is to never accept AI-written code without reading and understanding it. A developer who skips the review step gradually loses the ability to debug and maintain, until they depend on the AI to explain code the AI itself wrote.
The second principle is to ask the AI questions about the code it produced. "Why did you choose this approach?" and "What are the drawbacks of this solution?" are questions that create genuine cognitive dialogue and allow the developer to stay in control of the direction.
The Developer's Role in the Age of AI Coding Agents
The question Rosensweig hears most often is: "Will AI replace me?" His answer is research-based. A developer using AI coding agents produces five to ten times more code per day. But a developer who does not understand what the AI is doing produces five to ten times more problems.
The next generation of developers will not write code. They will manage code. They will define requirements, validate outputs, write tests, and ensure the agent understands broader context. This is not less cognitive work. It is cognitively different work.
How R&D Teams Prepare Correctly
Rosensweig offers a practical framework for teams starting out. The first stage is familiarization with tools through defined and limited tasks. The second is building an organizational review process for AI-generated code. The third is defining metrics that allow evaluation of whether the AI is raising output quality or merely output volume.
The most important point, in his view: AI does not replace thinking, it streamlines execution. An organization that hands its thinking to an AI agent will lose its knowledge advantage in a way that is not easy to recover.