Software coding agents have seen significant advancements since the introduction of CLI coding agents like Claude Code in early 2025. These agents excel in tasks such as writing complex code, debugging, and deploying applications. However, the article argues that coding agents are outliers in the AI landscape. Their success stems from characteristics unique to programming—deterministic outputs, clear success states, and rich tooling—which don’t apply to most other domains. The article highlights the unique feedback loop in the programming field, where those building AI models are also the primary users. This overlap leads to rapid improvements in coding agents. The abundance of training data related to programming, from code repositories to software development practices, further enhances their capabilities. In contrast, specialized fields like material science or drug discovery lack publicly available knowledge, making it harder to develop effective AI agents. A key distinction is made between workflows and agents. Workflows follow fixed sequences to achieve specific goals, while agents adapt and plan based on the problem at hand. While coding agents thrive in environments that require dynamic planning, many other domains would benefit more from structured workflows. The challenge lies in extracting and codifying the tacit knowledge of experts in those fields, which is often unarticulated and situational. The article cautions against assuming that the success of coding agents will translate to other domains, as their unique conditions—like immediate feedback and clear success metrics—are not universally applicable.