C · Agents and autonomy

Chapter 17. Autonomy Change and Its Detection

The rule that autonomy increases require re-authorization, its deliberate narrowing, and the unresolved tension that narrowing created.


17.1 The rule and its narrowing

The original rule was absolute: any autonomy increase is a Major change, always, re-entering risk classification and both G2 and G4.

That rule was narrowed during revision to apply to boundary-widening changes only — changes that expand what the agent may do, not changes within an already-approved boundary. The reasoning was practical: treating every configuration adjustment inside an approved envelope as a Major change made agentic systems ungovernable in practice, and a rule that cannot be followed is not followed.

The narrowing is logged as an unresolved trade-off, not a settled improvement. It reopens a gap that audit sampling is expected to cover, and the interaction between the two has never been tested. You are adopting a control whose designers explicitly flagged uncertainty about its calibration.

17.2 What counts as boundary-widening

Table 45.

Change Boundary-widening? Classification
New tool added to the allow-list Yes Major
Existing tool’s action limit raised Yes Major
Existing tool’s action limit lowered No Minor
Human-in-the-loop removed from a path Yes Major
Escalation threshold raised Yes Major
Escalation threshold lowered No Minor
New counterparty added to an approved list Yes Material to Major by blast radius
Agent invoked by an additional agent Yes, if composite authority grows Major
Model swapped inside an unchanged boundary No, on autonomy grounds Per model-change rules
Rate limit increased Yes, if it enables materially greater aggregate effect Material to Major

The asymmetry is deliberate. Reducing authority is safe by construction and should not carry governance friction; expanding it is the thing being controlled.

17.3 Making increases detectable

A prohibition whose violation cannot be detected is not a control. Three mechanisms make autonomy expansion visible.

Machine-readable boundaries at G3. The recorded boundary must be in a form comparable against live IAM and tool configuration. Without this, detection depends on someone noticing.

Continuous comparison. The Control Plane compares an agent’s live permission set against its G2-approved boundary. Divergence is a Security or Configuration drift signal with a specific, high-value meaning.

Verification latency ceiling. Post-action verification must occur within a defined maximum interval, scaled to tier. For a Tier 4 agent acting continuously, a weekly reconciliation is not verification.

[Practice recommendation] Treat any detected divergence between an agent’s live permissions and its approved boundary as an incident rather than a drift finding, regardless of whether the expansion looks benign. The alternative — triaging each case on merit — is how boundaries erode, because each individual expansion is usually defensible.

17.4 What remains unsolved

Two gaps are worth holding in view as you operate agentic systems.

Cumulative exposure. Many sub-threshold actions accumulating between verification cycles are not governed (Chapter 13, §13.4). Compensate with aggregate limits.

The narrowed rule’s interaction with sampling. Whether boundary-widening-only classification, backed by audit sampling, catches as much as the original absolute rule is unknown. [Practice recommendation] Until evidence exists, weight your Tier 3–4 audit sample deliberately toward agentic systems and toward changes classified as within-boundary. If that sample keeps finding misclassified expansions, revert locally to the absolute rule for agents and record why.