The first AI agents were built around a simple loop: task, LLM, tools, result, repeat.
For prototypes that is enough. But once an agent runs for a long time, interacts with other agents, changes the state of systems, spends resources, and performs real actions, a single loop is no longer enough.
A separate architecture grows around the model.
Intent turns into specification.
Tools and devices are described as capabilities.
Constraints and budgets become policy.
Execution produces evidence.
The result goes through validation.
State becomes reproducible and verifiable.
In such a system the LLM is no longer the whole system. It is one computational component inside a managed runtime.
That is exactly why Architecture-as-Code, Specification-as-Code, Policy-as-Code, Harness-as-Code and Evals-as-Code are now beginning to converge.
The next important question is no longer only how smart the agent is.
What matters far more: can you precisely define what it is allowed to do, what state it may change, how many resources it may spend, what evidence of the result it must provide, and where the boundary of autonomy lies.
To me, this is exactly the next stage of agentic architecture: not just more autonomy, but more governability, verifiability, and architectural discipline.