OPAL: Outcome-Predictive Agent Layer — a CNCF Sandbox project built on kagent

[HarshalSant]

Introducing OPAL — built on top of kagent

Hi kagent community 👋

I wanted to introduce OPAL (Outcome-Predictive Agent Layer) — an open-source Kubernetes control plane extension I have built using kagent as its agent runtime and A2A foundation.

Repo: https://github.com/opalnewharshal/opal

---

What is OPAL?

OPAL introduces two capabilities that have never existed in the cloud native ecosystem:

1. Outcome-Native Kubernetes API

Instead of writing YAML configuration, operators declare business outcomes:

apiVersion: opal.io/v1alpha1
kind: WorkloadOutcome
spec:
  slo:
    latencyP99Ms: 80
    availability: "99.95"
  cost:
    monthlyCeiling: "400"
  compliance: [PCI-DSS]

OPAL derives and owns all Kubernetes configuration (resources, replicas, HPA, PDB) required to achieve the declared outcomes.

2. Predictive Reconciliation

Every existing Kubernetes controller is reactive — it fixes drift after it occurs. OPAL introduces ClusterForecast (the first Kubernetes resource designed to hold future state predictions) and acts before drift happens.

---

How OPAL uses kagent

OPAL ships four kagent agents that form the AI control plane, communicating via A2A:

Agent	Role
ForecastAgent	Predicts future cluster state every 15 minutes
ReconcilerAgent	Acts on predictions before problems occur (the novel one)
OutcomeAgent	Translates declared outcomes → k8s config
AuditAgent	Every AI decision becomes a Kubernetes Event

kagent solves "how do AI agents live in Kubernetes". OPAL solves "what should those agents actually do" — the part nobody has built yet.

OPAL ships kagent as a Helm dependency and all OPAL agents are standard kagent Agent CRDs.

---

Why I am sharing this here

1. Feedback — I would love input from the kagent community on the agent design, especially the A2A communication pattern between ForecastAgent and ReconcilerAgent.

2. Collaboration — OPAL depends on kagent. Any improvements to kagent (persistent agent loops, better A2A routing, tool composability) directly benefit OPAL.

3. CNCF Sandbox — I am planning to submit OPAL for CNCF Sandbox. Given that OPAL builds directly on kagent, I wanted to introduce the project to the kagent maintainers first and explore whether there is interest in co-presenting or co-sponsoring the submission.

---

Links

• GitHub: https://github.com/opalnewharshal/opal
• Architecture: https://github.com/opalnewharshal/opal/blob/main/docs/architecture.md
• Getting Started: https://github.com/opalnewharshal/opal/blob/main/docs/getting-started.md

Looking forward to the conversation!

[HarshalSant]

Update: The OPAL repository has been transferred to the opalnewharshal GitHub organisation.

New canonical URL: https://github.com/opalnewharshal/opal

CNCF Sandbox application: cncf/sandbox#492

Looking forward to any feedback from the kagent community!