Decision Rights for AI Agents

The documented authority of an AI agent to take a class of action without further human approval — specifying who approved that authority, for what scope, under what justification, and for how long. The formulation comes from Knostic’s framing in AI Coding Agent Governance (Knostic, 2025–2026): “Governance refers to defining who has the authority to act and under what justification.”

Decision rights are the governance counterpart to least privilege. Where least privilege answers “what can the agent access?”, decision rights answer “what can the agent decide on its own — and on whose authority?”

Why it’s distinct from access scoping

Access scoping (D3 of Agentic AI Security Capability Maturity Model — A 2026 Practical Proposal) controls reach: which repos, which APIs, which environments. Decision rights control authority delegation within that reach: among the actions the agent can take, which can it take autonomously, which require notification, which require approval, which are prohibited?

ConceptQuestion it answersCMM domain
Least privilegeWhat can the agent access?D2 + D3
Least agency (Least Agency Principle)How autonomous is it?D3
Decision rightsOn whose authority + under what justification?D1 (governance) + D3 (control)
AuditabilityWas the decision reversible and attributable?D7 + D9

Without decision rights, an agent operates with apparent authority — the user delegating to it has implicitly granted it the authority to act in their name, but the delegation chain is undocumented. This is the failure mode that turns helpful agents into risk multipliers.

A decision-rights matrix (sample)

A decision-rights matrix per agent type / agent-role pair is the artifact that operationalizes this concept. Sample row:

Agent typeAction classDecision rightApproverJustificationTime bound
code-review-agentcomment on PRautonomousn/atask-definedalways
code-review-agentmerge approvalrequires approvalnamed human reviewerrisk-tier=medper-PR
deployment-agentdeploy to stagingautonomousn/aCI greenmaintenance window only
deployment-agentdeploy to productionrequires approvalrelease manager + on-callchange ticketper-deploy
deployment-agentforce-push or branch deletionprohibited

A row is not optional for any action class an agent can take. “We didn’t write a row for this” defaults to prohibited.

Relation to OWASP Least Agency tiers

Maps directly to the OWASP four-tier least-agency model documented in Least Agency Principle:

OWASP tierDecision-right wording
autoautonomous
notifyautonomous-with-notification
confirmrequires approval
blockprohibited

The OWASP tiering is the mechanism; decision rights are the justification record — who approved the tier assignment, under what risk analysis, with what review cadence.

Place in the CMM

  • D1 L3 evidence: documented decision-rights matrix per registered agent type (sharpening from the Knostic ingest).
  • D3 L3+ evidence: the matrix is operationalized in the PDP (Cedar / OPA policies) — every tool call is mediated by the matrix at runtime.
  • D9 L4 evidence: HITL approval-rate metrics per matrix row (rubber-stamp detection).

Halt-authority is also a decision right

Distributed Kill Switch — the practice of putting halt-authority in the hands of every team member in the loop — is itself a decision right that must appear in the matrix. Sample row:

Agent typeAction classDecision rightApproverJustificationTime bound
any-agenthalt current runautonomous (any in-loop human)n/asafety; logged-not-punishedalways
any-agenthalt workflow / use caseautonomous (workflow owner)n/asafety + outcomeper-use-case
any-agenthalt project / line of businessrequires approvalCouncil (Legal + COO seats)escalationper-project

Without explicit halt rows, the practice has no audit trail and quietly degrades to “the CIO holds the kill switch alone” — the failure mode the talk explicitly warns against.

Relations