Enterprise DecisionOps Control Plane

The Enterprise DecisionOps Control Plane demonstrates proof-carrying operations for Agentic AI.

It proves why an agent acted by generating a verifiable evidence chain mapped to NIST AI RMF outcomes. It bridges the gap between autonomous execution and enterprise compliance.

📦 Download Sample Audit Packet

Extract this zip to see exactly what “cryptographic-style evidence” looks like under the hood: OPA policies, SBOMs, trace JSONs, and evidence appended in a real AI invocation.

The Architecture

The architecture enforces that no agent action happens without a rigorous, context-aware policy decision:

flowchart LR
    Agent[AI Agent\nvia MCP] --> Gateway(Gateway Manager\nOPA Engine)
    Gateway --> Trace[Decision Trace Event]
    Trace --> Broker((Stream\nKafka))
    Broker --> Worker[Worker Action]
    Worker --> Evidence[Evidence Append]
    Evidence --> OTel(Observability\nJaeger)
    OTel --> Audit[Immutable Audit Packet]

1. Trace Evidence (The “Audit Packet”)

Agents request actions via the Model Context Protocol (MCP), passing execution attributes (e.g., tests_passed). Workflows track these via a correlation_id to output an immutable Audit Packet.

The Decision Trace (decision_traces.jsonl)

Generated by the Gateway. Proves that OPA semantically inspected the AI’s execution context and authorized the action based on a specific bundle_hash.

{
  "event_id": "3f2bd2e56...",
  "correlation_id": "corr-500bdcb7dacac332",
  "request": {
    "action": "deploy",
    "environment": "dev",
    "attributes": {
      "tests_passed": "true"
    }
  },
  "policy": {
    "engine": "opa",
    "bundle_hash": "sha256:84e67e166187a795f5...",
    "decision": "allow",
    "rationale": [
      {
        "code": "ALLOW_DEPLOY_DEV",
        "message": "Dev deploy permitted because tests passed."
      }
    ]
  }
}

The Evidence Append (evidence.jsonl)

Generated by the Worker. Proves that the action allowed by the policy was successfully executed by the target workload, closing the loop.

{
  "event_id": "69fc7258...",
  "correlation_id": "corr-500bdcb7dacac332",
  "evidence_type": "action.executed",
  "action": "deploy",
  "result": "ok"
}

2. Observability & OpenTelemetry Parity

Standard logging fails for async AI pipelines. This platform implements OpenTelemetry (OTLP) natively.

Using correlation_id for W3C Trace context, operators can drop into Jaeger to see exactly where an AI action was blocked or allowed across network boundaries.

3. Cryptographic Software Supply Chain (SLSA-lite)

How do we inherently trust the gateway and worker binaries generating the evidence? By verifying their origins.

The architecture includes an automated supply chain pipeline:

• SBOM Generation: Uses Syft to extract an exhaustive SPDX JSON bill of materials for all executing artifacts.
• Image Signing: Uses Cosign to cryptographically sign the images and attest the SBOM.
• Verification: The Audit Packet export packages these artifacts so that reviewers can verify the physical binaries executing the rules were free of critical CVEs at the time of the AI decision.

The Bottom Line

By decoupling execution intent from authorization, this Control Plane lets autonomous systems operate safely while outputting the exact cryptographic evidence enterprise governance demands.