Andrej Karpathy and the new value of engineering judgment
As coding agents take on more implementation, the scarce skill shifts toward context, verification, and understanding what the software should do.
Editorial image for Andrej Karpathy and the new value of engineering judgmentVIBECIO Editorial Desk ·
Andrej Karpathy’s public account of the agentic shift puts a different question in front of engineering leaders. If increasingly capable agents can perform substantial implementation work, what should the human engineer become better at? His answer emphasizes understanding, the context supplied to the model, and the feedback used to judge its work.
In his Sequoia Ascent 2026 post, Karpathy describes programming through prompts, context, tools, and instructions. He distinguishes the accessibility of vibe coding from the discipline of engineering with agents. The post contains an AI-generated summary and cleaned transcript that he says he reviewed; it is a useful first-party record of his public thinking, not an interview conducted by this magazine.
Look at the feedback loop
Karpathy’s autoresearch repository makes part of that thinking concrete. An agent modifies a bounded training file, runs an experiment within a fixed time budget, evaluates a defined metric, and keeps or discards the change. The human works on the instructions that shape the research process. The project is deliberately small, with a constrained edit surface and a visible record of experiments.
That design is instructive for enterprise teams even when their problem has nothing to do with training a model. Progress becomes easier to assess when the task has a known boundary and a credible way to distinguish an improvement from a regression. The difficult work is often defining that evaluation before the agent begins optimizing against it.
A passing test can establish only what the test checks. If a team asks for a faster service but ignores access control, recoverability, or the meaning of an error response, it has described an incomplete objective. An agent can satisfy the visible target while leaving the system less useful. Better verification begins with a better account of what matters.
Understanding stays with the team
Karpathy’s discussion connects stronger agents with the continuing need for human understanding. VIBECIO’s interpretation is that engineering judgment becomes more valuable when implementation is easier to obtain. Someone still has to decide whether a requirement is coherent, a dependency is understood, and a proposed result deserves further investigation.
This creates an opportunity for a more deliberate apprenticeship. A newer engineer can use an agent to trace a request, compare implementations, and propose tests. An experienced colleague can examine the explanation with them, identify missing assumptions, and connect it to operational history. The learning comes from interrogating the result together and observing what happens when the assumptions are tested.
The relationship is reciprocal. People who are comfortable exploring with agents may discover faster ways to investigate a system. Engineers who have maintained services through failures can explain where a convincing implementation may still be fragile. Neither perspective has to diminish the other for the team to improve.
Generated code still enters a real system
Once an agent can use tools, the work extends beyond text generation. NIST’s tool-use discussion separates capabilities through dimensions such as access, monitoring, and reversibility. Those distinctions belong in an engineering workflow: reading code, staging a change, approving it, and deploying it are different powers.
The practical opportunity is to make exploration more ambitious while keeping decisions legible. Give agents a useful task, a bounded environment, and feedback that reflects the actual requirement. Preserve the team’s ability to explain the result. The valuable engineer is increasingly the person who can connect the speed of the tools to an understanding of the system they are changing.
Sources