Autonomous agents need more than guardrails. They need governance.
A security control plane for agent identity, delegated authority, policy enforcement and accountable execution.
Deny by default
Every tool call is evaluated. Unknown and high-risk tools are denied even when no policy matches.
Agent identity
Every agent gets a non-human identity with a hashed, rotatable, revocable credential.
Verifiable evidence
Decisions land in a hash chain you can export so a third party can recompute them.
Execution assurance
Commit Guard and independent execution verification are specified, not yet shipped.
Guardrails are not governance.
A guardrail stops one thing. Governance answers the questions an auditor asks:
- Who is acting? Not a service account shared by three workflows.
- Under what authority did it act, and was that authority still valid at the moment of the call?
- What did it decide, and on what basis?
- Can a third party independently recompute that decision six months later?
Most agent frameworks answer the first question badly and the last three not at all. AWARE exists to answer all four, in that order.
The chain of authority
Each stage is a place where an agent’s authority can be checked, narrowed or cut. AWARE implements the highlighted stages today.
The control plane, component by component
Each capability below is labelled with its implementation state, taken from the AWARE capability register. Where something is not shipped, this site says so.
Agent identity
UUID identity, PBKDF2-hashed credentials, timing-safe comparison, rotation, and session tokens with trust-domain claims.
Read more →Policy engine
ALLOW, DENY, AUDIT or NO_POLICY for every tool call, evaluated against tool risk, data tier and the policy store. Hot reload, no restart.
Read more →Revocation
Revoking an agent invalidates its credential immediately and writes the decision to the audit trail.
Read more →Audit & evidence
A hash chain with an independent verifier, exportable windows a third party can recompute, and failure outcomes recorded.
Read more →Compliance mapping
Seven frameworks, 184 of 243 CSA AICM v1 controls across all 18 domains, and MITRE ATLAS technique mapping.
Read more →T0-T4 model
Five tiers that classify and promote an agent’s autonomy posture, with validated promotion events.
Read more →Read the source, not the pitch
AWARE is Apache-2.0 and the capability register that governs this site is public. If we claim something, there is a file and line number behind it.