agent · AVC

Capabilities are explicit contracts.

Inputs, outputs, authority, risk, evidence and failure behavior are declared before use.

CapabilityAuthorityRiskEvidence
research.executeExecute reversibleR1Research receipt
outreach.preparePrepare onlyR2Draft + rationale
outreach.sendApproved executionR3Approval + delivery receipt
policy.modifyForbiddenR4Blocked event
Capability anatomy

Each grant answers seven operational questions.

Missing answers cause denial rather than best-effort guessing.

FieldQuestionExample boundary
SubjectWho may act?Named agent or workload identity
PurposeWhy is the grant issued?One mission or product outcome
ScopeWhich resources and actions?Allowlisted repository, account or adapter
EnvironmentWhere may it run?Local, test or approved production target
RiskWhat is the ceiling?R0 through R3; ordinary grants exclude R4
TimeWhen does it expire?Bounded validity window
EvidenceHow is use reviewed?Authorization, action and outcome receipts
Grant model

A capability is purpose-, scope-, environment- and time-bound.

The grant names the subject, resources, actions, risk ceiling, validity and revocation behavior.

Risk ceiling

A capability cannot authorize work above its declared risk.

R3 work requires independent approval, while R4 remains outside ordinary capability grants.

Revocation

Authority can be narrowed or withdrawn while a mission exists.

The runtime must re-evaluate current authority before consequential transitions rather than assuming an earlier grant lasts forever.