Standalone Policy Engines (OPA, Cedar, Cerbos) vs. Full-Lifecycle Action Assurance
Why returning 'Allow' or 'Deny' is only 15% of the engineering work required to safely govern autonomous ERP and financial mutations.
Key Takeaways
- OPA, Cedar, and Cerbos are fast stateless policy evaluators that answer binary Allow/Deny queries.
- High-risk agent actions require a third decision—RequireApproval(N)—plus stateful approval queues, SoD checks, idempotency locks, and ERP read-back.
- Building those surrounding 85% of controls atop a raw policy engine takes 6–12 weeks of custom distributed systems engineering.
1. Business & threat context
When platform engineers first scope an agent control layer, a common instinct is to drop Open Policy Agent (OPA) or Cerbos into a sidecar and write Rego rules. For stateless microservice read/write permissions, that pattern works well.
As soon as the CFO asks 'What happens when an invoice is over €5,000?', a stateless policy engine hits a wall. Returning 'RequireApproval' means you now have to build a persistent state machine, an OIDC-authenticated approver inbox with WebAuthn step-up, RFC 8693 self-approval prevention, canonical payload diffing, short execution leases, timeout reconciliation, and a hash-chained audit vault.
Save 8+ weeks of plumbing
Don't spend a quarter building approval state machines, replay locks, and audit crypto around a stateless policy sidecar.
2. How ActEscrow solves it (Service 01)
ActEscrow includes a deterministic, versioned policy engine out of the box—and pairs it with the complete stateful lifecycle from pending approval through downstream ERP read-back verification and ECDSA evidence anchoring.
{
"states": [
"PendingApproval -> Approved -> Executing -> Executed (Verified)",
"PendingApproval -> Rejected | Canceled | Expired",
"Executing -> Unknown (Timeout) -> Reconciled (Executed | Failed)"
]
}3. Production implementation checklist
When evaluating Open Policy Agent (OPA) / Cedar / Cerbos alongside a customer-owned enforcement boundary, engineering and finance leaders should verify the following operational controls:
- Support three-way policy decisions: Allow, Deny, and RequireApproval(N)
- Bind approvals to canonical SHA-256 digests so approved parameters cannot drift
- Reconcile ambiguous downstream timeouts before permitting any retry
Going deeper: what a stateless engine can and cannot hold
To be fair to the tools named above: OPA can return structured documents rather than a bare boolean, and Cedar and Cerbos are fast, well-designed evaluators. You can model a RequireApproval result in any of them. What none of them holds, by design, is the state that follows that result: who has voted, whether those votes bind to the same bytes, whether the action was already redeemed, and what the ERP said afterwards.
The effort figures earlier in this post, such as percentages of work and week counts, are rough planning assumptions rather than measurements. The shape of the work is the reliable part: an approval state machine, SoD rules, an execution lease, timeout reconciliation and tamper-evident evidence all sit outside the evaluator.
If you already run OPA or Cedar
You do not need to remove a policy engine that works well for service authorization. A reasonable split keeps OPA or Cedar for coarse questions, such as whether this workload may call the gateway at all, and lets the gateway's deterministic engine decide action-level outcomes with its versioned, maker-checker-governed policy. The important rule is that only one component decides each question, so there is never doubt about which policy version applied.
- Keep existing engines for service-to-service and API-level authorization.
- Use action-level policy where approvals, digests and evidence are needed.
- Do not evaluate the same action in two engines with different rules.
- Test with the simulate endpoint before activating a new action policy version.
Properties to require of any action-level evaluator
Whichever engine decides action outcomes, require the same properties: default deny, ordered and versioned rules, no clock, network or model calls during evaluation, missing or mistyped parameters never satisfying a condition, and a dry-run mode that shows which rule would match. Those properties are what make a decision reproducible during an audit.
Security scope and deployment prerequisites
This is a pilot implementation, not a zero-risk guarantee or compliance certification. Examples describe repository capabilities or deployment recommendations, not proof that a live installation has enabled them.
Every enrolled write fails closed during control-plane outages. Privileged control-plane MFA is mandatory outside Development, but accepted MFA claims do not prove recent authentication or hardware-backed WebAuthn unless the identity provider enforces those properties.
Per-agent deactivation blocks new evaluations and redeems and cancels Pending/Approved actions; it cannot recall an already executing ERP write. Evidence verifies recorded events, not unrecorded ERP activity; independently retained heads and trusted keys are needed to detect truncation or replacement.
Production blockers remain
The static portal still uses sessionStorage tokens (BFF not implemented). Off-host backup signing, immutable backup storage, live IdP MFA/logout testing, restore drills, adapter validation, monitoring, and legal review require deployment-specific work. Anchoring and SIEM forwarding are optional and must be configured and tested. See the repository deployment guide and runbook gap register before using real customer data.
Related deep dives
GRC Automation (Vanta, Drata) vs. Cryptographic Decision Evidence Vaults
Nov 20, 2026 · 3 min readArchitectureiPaaS Connectors (Workato, Zapier, Make) vs. Fail-Closed Agent Gateways
Nov 19, 2026 · 3 min readArchitectureLLM Observability (LangSmith, Arize, Galileo) vs. Pre-Execution Control
Nov 17, 2026 · 3 min read