​Will Physical AI Bring A “Coding Agent” To Physical Businesses?

Maria Zhang is the CEO and Co-Founder of Palona.

Coding agents changed software development. A parallel shift in physical businesses requires models that understand real-world interactions, not just what appears in a scene.

Coding agents can investigate projects, implement changes, run tests and revise their work. Their significance goes beyond speed. They change what developers can delegate and where human judgment matters.

​What would an equivalent capability make possible for a restaurant, store or warehouse? Rather than writing software, such an agent could understand the business, help people improve it and evaluate the results.

​Building that capability requires more than connecting an agent to systems or cameras. Physical businesses need a different technical foundation.​

Software Gives Agents A Head Start

OpenAI’s “Harness engineering: leveraging Codex in an agent-first world” describes an internal beta whose code was generated entirely by Codex. Engineers defined requirements, designed the environment and built feedback loops. Their work changed; it did not disappear.

​The team made documentation, application behavior, logs and metrics accessible to agents. This illustrates the broader requirement that agents need environments they can inspect, act within and learn from. As Anthropic’s engineering guidance explains, code and automated tests support iterative workflows, although tests cannot verify every business requirement.

​Physical businesses lack comparable visibility. Order timestamps cannot explain whether preparation was slow or a finished item waited unnoticed. Transaction systems expose records, not the activity between them. Existing cameras capture that activity, but footage alone does not explain the workflow.​

The Camera Sees People; The Operator Sees A Queue

​The objective is continuous understanding of meaningful changes, rather than answers to occasional questions about a scene.​