“Agentic AI” is often discussed as if it named a stable kind of system. In practice, systems described as agentic vary widely. Some perform a short sequence of tool calls. Others can continue acting, change plans, coordinate components or operate across several organisational systems.
The label alone says little about what a particular system can do or how it is governed. A more useful analysis examines the choices that shape its practical agency.
Configuration
Configuration establishes the system’s working conditions before it acts. This includes its goals, instructions, permissions, available tools, accessible information and limits. It also includes who can change those settings and whether changes remain visible.
These choices shape the system’s mandate. Two systems built with the same underlying model may behave very differently because one can only retrieve documents while another can send messages, alter records or initiate transactions.
Runtime
Runtime concerns what happens while the system is operating. How is it activated? Can it decide what action to take next? How long may it continue? Can it retry after failure, create sub-tasks, call other agents or change its route when conditions differ from the original plan?
Continuation is particularly important. A system that completes one bounded action creates a different governance problem from one that keeps working until it judges a goal complete. The ability to continue, branch or recover gives the system practical latitude even when its high-level objective was set by a person.
Supervision
Supervision covers how people or other systems observe, evaluate and control what happens. It includes logging, review points, approvals, alerts, intervention routes and termination mechanisms.
Supervision should not be reduced to the presence of a person. What matters is whether the supervisor receives useful information in time and has the authority and practical means to change the course of action.
A design grammar
Looking across configuration, runtime and supervision provides a design grammar for comparing agentic systems. It shifts attention from whether a product is “really agentic” to more concrete questions:
- What activates the system?
- What mandate and permissions does it receive?
- What choices can it make while operating?
- What allows it to continue or requires it to stop?
- What can supervisors observe?
- Who can interrupt, redirect or terminate it?
This approach also makes research easier to organise. Studies that appear to concern the same category of AI may actually examine different configurations, levels of autonomy or supervisory arrangements.
Why the distinctions matter
Broad claims about agentic AI can obscure where risks and responsibilities arise. The consequential feature may not be the model’s apparent intelligence. It may be a permission, a trigger, a missing stopping condition or an oversight role that exists only on an organisational chart.
Governance becomes more precise when it addresses those design choices directly. It can then ask whether the system’s practical freedom is matched by sufficient observability, intervention capacity and accountability.