Should coding agents be treated as constrained executors rather than architectural authorities?

I work independently and use ChatGPT and coding agents to help me design and build software systems. As my projects became larger, I noticed that the main problem was no longer getting the AI to generate code. The harder problem was retaining control over what the agent was allowed to change, how it interpreted the architecture, and what counted as completed work. This led me to a working principle: A coding agent can be a capable executor, but it should not become the architectural authority of the system. At a high level, I now try to separate several responsibilities: The human defines the system’s intent, architecture, boundaries and non-negotiable constraints. The agent receives narrowly defined units of work rather than unrestricted authority over the repository. Project state is maintained outside the conversation so that a new session does not have to reconstruct the system from memory. An agent’s statement that a task is complete is treated as a claim, not as evidence. Tests, repository state and explicit acceptance are used to determine whether work is actually complete. Progression to the next stage requires a deliberate decision rather than an assumption by the agent. Unclear or unauthorized actions should fail closed instead of being interpreted creatively. I am deliberately not describing the operational implementation because I am still evaluating the underlying reasoning. I would like criticism from people who have built or supervised real agent workflows: Is this a legitimate governance problem, or am I converting a coding workflow into unnecessary bureaucracy? Which controls become essential once agents can modify files, execute tools and make multi-step decisions? Where should human authority end and agent autonomy begin? What failure modes would this model still fail to prevent? Does an established discipline already cover this combination of authority, scoped execution, external state, verification and explicit acceptance? I am not selling a tool and I am not looking for recommendations about which coding agent to use. I am trying to determine whether this model of controlled execution is technically sound.