The control plane
AWARE is a set of enforcement layers that sit between an autonomous agent and the tools it calls. Each layer narrows what the agent can do, and each layer records what it decided.
Agent identity
A non-human identity per agent, with a credential that is stored only as a salted hash, compared in constant time, rotatable, and revocable. Session tokens carry trust domain, blast radius and execution context.
Policy engine
Every tool call is evaluated before it runs. Unknown tools and high-risk tools are denied even with no matching policy. Decisions are ALLOW, DENY, AUDIT or NO_POLICY, with a stated reason.
T0-T4 constraints
Five tiers classify an agent’s autonomy posture and are promoted through validated events. The engine enforces by tool risk and data classification; the tiers record the posture.
Revocation
Revoking an agent flips its state so its credentials stop verifying, applies a blast-radius penalty to its trail, and writes the decision to the audit trail.
Audit & evidence
Decisions are written to a hash chain. An independent verifier can recompute it, and an exportable window lets a third party verify a slice without access to your database.
Execution assurance
Authorising an action and verifying it happened are different things. AWARE does the first. The second — atomic commit, target-side enforcement and independent execution verification — is specified but not built.