# Agent Workload Maturity Tiers

An actionable progression checklist for advancing agent workloads through Mandrel's four operational maturity tiers.

Source: https://mandrel-web-363efb.gitlab.io/docs/overview/workload-maturity/

---


This guide defines the actionable technical criteria required to advance an agent workload through Mandrel's four operational maturity tiers.

> **Maturity Alignment:** While **[Enterprise Governance Maturity](https://mandrel-web-363efb.gitlab.io/docs/white-papers/enterprise-governance-maturity.md)** measures an organization's overall program maturity from Shadow AI to a unified mesh, **Workload Maturity Tiers** govern the runtime promotion criteria for individual agent deployments. A mature organization enforces these automated workload gates within its CI/CD and runtime infrastructure.

---

## Tier Summary

| Tier | Posture | Identity & Access | Governance Gate | Target Capabilities |
| :--- | :--- | :--- | :--- | :--- |
| **Tier 1: Sandbox** | Private / Prototype | Anonymous / Dev Token | Static schema check | Read-only MCP resources |
| **Tier 2: Trusted** | Verified Workload | SPIFFE SVID + OIDC | `mandrel-cli validate` | Mandatory HITL (`challenge-always`) |
| **Tier 3: Federated** | Discoverable Mesh | Signed Realized Spec | Metrology Lab promotion | Intent classification & capability discovery |
| **Tier 4: Autonomous** | Hardened Runtime | SPIFFE Federation + DPoP | Continuous Conformance | Dynamic Client Registration & CAE revocation |

---

## Tier 1: Sandbox (Private Prototyping)

The entry point for rapid agent prototyping and local development.

### Progression Checklist
- [x] Author a baseline `MandrelSpec` declaring metadata name and namespace.
- [x] Scope agent tool calls exclusively to read-only MCP resources.
- [x] Run locally using Podman or Docker Compose behind a Mandrel Collet sidecar.

```yaml
apiVersion: mandrelproject.ai/v1
kind: MandrelSpec
metadata:
  name: sandbox-agent
  namespace: default
spec:
  governance:
    risk-profile: "balanced"
```

---

## Tier 2: Trusted (Verified Workload)

Transition from unverified prototype to a monitored workload operating within a strict governance envelope.

### Progression Checklist
- [x] Pass static validation without errors: `mandrel-cli validate agent-spec.yaml`.
- [x] Provision a workload SPIFFE SVID via SPIRE or the Gantry Operator.
- [x] Enforce mandatory Human-in-the-Loop (HITL) gates (`action: challenge-always`) on all high-risk or state-changing tools.
- [x] Define explicit delegation depth limits (`spec.runtime.max-hops-allowed`).

```yaml
spec:
  governance:
    risk-profile: "safety-first"
    enforcement-matrices:
      - type: "threshold"
        attribute: "transaction.value"
        tiers:
          - limit: 100.00
            action: "challenge"
  runtime:
    max-hops-allowed: 2
```

---

## Tier 3: Federated (Discoverable Mesh)

Workloads promoted to the governed agent mesh, enabling capability-based discovery and cross-agent delegation.

### Progression Checklist
- [x] Submit spec to the Metrology Lab promotion gate (`mandrel-cli agents promote`).
- [x] Obtain an approved, Ed25519-signed **Realized Spec** with platform policy ceilings baked in.
- [x] Configure Layer 0 **Isolated Intent Verification** (`spec.governance.intent-verification`).
- [x] Register declared capabilities in `spec.capabilities.intents` for late-binding discovery by orchestrators.

```yaml
spec:
  governance:
    intent-verification:
      mode: "strict"
      failure-action: "deny"
  orchestrator:
    permitted-intents:
      - name: "Execute_Payout"
        description: "Fund transfer intent routing."
```

---

## Tier 4: Autonomous (Scalable Production)

The highest assurance posture, enabling autonomous execution bounded by real-time revocation and cryptographic proof of possession.

### Progression Checklist
- [x] Enforce sender-constrained tokens via **DPoP (RFC 9449)** (`spec.identity.enforce-dpop: true`).
- [x] Subscribe the Collet sidecar to **Continuous Access Evaluation (CAE / RFC 8936)** for sub-second token revocation.
- [x] Pass Metrology Lab **Golden Dataset Benchmarking** during automated re-promotion pipelines.
- [x] Enable Dynamic Client Registration (DCR) for automated SPIFFE trust domain federation.

---

## Promotion Workflow

```mermaid
graph LR
    T1[Tier 1: Sandbox] -->|mandrel-cli validate| T2[Tier 2: Trusted]
    T2 -->|Metrology Lab Promotion| T3[Tier 3: Federated]
    T3 -->|CAE + DPoP + Benchmarks| T4[Tier 4: Autonomous]
```