Action proposed
The agent proposes an action with its target, parameters, scope and asserted intent.
WitnessOS is not an operating system for machines. It is the operating layer for governed agent action, the kernel between "the agent wants to" and "the agent may".
Deterministic policy, human approval, tamper-evident receipts and independent timestamps. Auditor-verifiable with zero trust in us.
AI governance is moving from guidance to evidence. Audit, risk and insurance teams will ask organisations to prove what autonomous systems actually did, not simply show what a platform says happened.
"Show me, don't tell me."
When an agent spends money, sends a message or touches a governed action, who can prove what happened, who approved it and when? Self-attested logs that a vendor could edit are not enough.
The agent proposes an action with its target, parameters, scope and asserted intent.
Deterministic, OPA/Rego-compatible policy returns Allow, Deny, Conditional or Escalate.
The evaluation produces a tamper-evident receipt with the action hash, policy version, context, decision, timestamp and chain of custody.
Execution proceeds only when sanctioned. Denied actions receive structured remediation from advisory to failsafe, with every escalation recorded.
The public ladder closes at E4. Every grade describes the evidence a receipt can support, and unprovable gates cap the grade down, never up.
POLICY
OPA-native Rego policy-as-code, evaluated deterministically before action.
EVIDENCE
SHA-256 signed, tamper-evident and exportable as audit bundles.
REMEDIATION
Structured escalation from advisory guidance to hard failsafe.
BINDING
TPM or fingerprint binding for governed licensing.
INTEGRATION
Designed for REST, WebSocket and SDK integration.
CONTROL
System, environment and action-level policy tiers.
Strict mode issues evidence anchored by a real third-party RFC 3161 authority.
Altered records cannot reach E4.
Unsanctioned action types fail closed.
Evidence degradation cannot drift silently.
Including real-CA timestamp verification.
WitnessOS is the runtime governance layer for autonomous agents, developed by Empire Labs Pty Ltd. It sits between an agent wanting to act and the agent being allowed to act, applying deterministic policy, human approval and tamper-evident evidence to every governed action.
An agent transaction is any governed action an autonomous agent performs, such as a payment, refund or message. WitnessOS records each one as a signed, chained evidence receipt with an independent timestamp, so the transaction can be verified later without trusting the operator.
WitnessOS governs payment and commerce actions through connectors, including x402 payments, Stripe refunds and messaging. Evidence receipts are produced before a connector runs, giving AI commerce the audit trail it needs: what was spent, who approved it, and proof the record was not altered.
E4 is the top grade on the public WitnessOS evidence ladder. It means the receipt was independently timestamped by an RFC 3161 Time Stamping Authority with revocation checked at record time, and it re-verifies offline with zero trust in Empire Labs.
No. WitnessOS is not an operating system for machines. It is the operating layer for governed agent action, the kernel between "the agent wants to" and "the agent may". It governs, approves and proves actions rather than running the agent itself.
Empire Labs Pty Ltd is a participant in Agentic AI Foundation working groups on observability, traceability and agentic commerce. WitnessOS evidence concepts and reference implementations are shared in those groups as open contributions. AAIF membership is held by the foundation's member organisations; Empire Labs is not an AAIF member org.
Book a 30-minute technical assessment to map governed actions, policy gates and evidence requirements for your environment.