Agency and control¶
Agency is the capacity to act toward a goal. Autonomy is the degree to which an agent can choose and modify its actions without external intervention. Control is the set of constraints that bound those actions.
The three are related but not identical. A system can be highly agentic and tightly controlled (a well-governed agent), or weakly agentic and poorly controlled (a script with broad credentials).
Control is not knowledge¶
A policy can be written down, described in a prompt, or represented in a model’s context. None of those enforce it. Enforcement requires an architectural mechanism: a policy decision point, a sandbox, a capability allowlist, a gateway, or an approval path.
This distinction is central to DIKCA: knowledge answers “what is the case,” control answers “what may be done.”
Meta-control¶
Autonomy at the highest level is meta-control: the capacity to modify one’s own control structures. An agent that can change its own policy, capabilities, or plan is exercising meta-control.
Meta-control is not a feature to enable casually. It requires:
- an architecture that bounds what can be changed;
- separation of duties;
- change evidence;
- re-verification after the change;
- a way to revoke the change.
A practical distinction¶
| Question | Concept |
|---|---|
| Can the system pursue a goal? | Agency |
| Can it choose its own steps? | Autonomy |
| What bounds those steps? | Control |
| Can it change the bounds? | Meta-control |
| Who is accountable? | Governance |
Why it matters¶
- Calling a system “autonomous” without naming the bounds is marketing, not architecture.
- Adding control does not always increase safety; poorly designed control can destabilize behavior (see UAD).
- The right question is not “how autonomous is it?” but “what can it do, what stops it, and what proves it?”