Developer argues agent frameworks and runtimes are distinct layers after auditing own codebase

A software developer at soit-ai has published a technical breakdown arguing that 'agent platform' is an umbrella term that conflates at least three separate concerns: the library developers write, the engine that runs it, and the control plane that governs it. Using their open-source repository on GitHub as a specimen, they show the core agent loop — planner, executor, and verifier — totals just 239 lines of code. By contrast, the runtime governance layer, including a policy gateway, spans 475 lines on a single port alone, handling permissions, cost attribution, secret boundaries, and approval interrupts. The executor class itself contains no business logic, deliberately delegating all calls to a port where execution rules are enforced separately from the loop. The author contends this separation is intentional and structural, not incidental, and uses line counts rather than diagrams to make the case.