---
title: "Agent Gateway, Arc and Relate: Who Owns the Agent Governance Layer?"
slug: "agent-governance-layer-google-agent-gateway-chargebee-arc-neo4j-relate"
author: "Blanca Amigot"
date: "2026-11-21 12:00:00"
category: "Premium"
sector: "enterprise"
topics: "agent governance, ai agent gateway, google agent gateway, envoy, mcp, a2a, chargebee arc, agentic finance, neo4j relate, agent memory, human-in-the-loop, enterprise ai agents, model-agnostic, ibl.ai"
summary: "In November 2026 Google took Agent Gateway cross-cloud, Chargebee launched Arc with a human approval gate for finance agents, and Neo4j open-sourced Relate for agent memory. Three layers of agent governance, and the question of who should own them."
banner: ""
thumbnail: ""
linkedin: |
  Three launches in eight days, three layers of the same stack: the infrastructure that governs AI agents in production.

  Nov 11: Google put Agent Gateway for cross-cloud into Limited Preview. Agents on EKS, AKS, OpenShift or bare metal get a verifiable identity, and their MCP and A2A traffic runs through local gateway proxies that enforce policy. Google says it builds on Envoy and Kubernetes.

  Nov 11: Neo4j released Relate, an open-source embedded graph database in Rust for agent memory, under Apache 2.0.

  Nov 18: Chargebee launched Arc for finance. Its agents prepare the work; every action is shown, cited and held for review before anything is committed.

  Gateway, memory, approval gate. Rent each from a different vendor and your agents' rules, recall and audit trail live in three places you don't control.

  On ibl.ai, the gateway, the memory layer and the audit logs run on your infrastructure, on any model, and you own all the code and the data.

  #iblai #AgenticAI #EnterpriseAI #AIGovernance #MCP #AIAgents
---

## The Short Answer

**In November 2026 Google, Chargebee and Neo4j each shipped one layer of agent governance: a gateway that polices agent traffic, an approval gate that holds agent actions for review, and a memory store for what agents learn. ibl.ai runs all three layers as one model-agnostic platform on your infrastructure, where you own all the code and the data, instead of renting them from three vendors.**

Google put [Agent Gateway for cross-cloud](https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes) into Limited Preview on November 11. Neo4j released [Relate](https://neo4j.com/press-releases/announcing-relate-by-neo4j/) the same day. Chargebee launched [Arc](https://witharc.ai) on November 18.

None of the three is a model or a chatbot. Each is a control: what an agent may call, what it may commit, and what it remembers. That is the layer enterprises now have to decide where to put.

## What did Google ship with Agent Gateway, and where does Envoy fit?

Google's [release notes](https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes) for November 11, 2026 extend Agent Gateway to CNCF-conformant Kubernetes clusters outside Google Cloud, naming Amazon EKS, Azure AKS, Red Hat OpenShift and bare-metal clusters.

Agents get verifiable identities, agents and MCP servers are registered in Google's Agent Registry, and their traffic is routed through local gateway proxies.

It is free during Limited Preview; connected products such as Model Armor, Cloud Logging and Cloud Trace bill at standard rates.

A day later, Agent Gateway for GKE entered Preview for clusters on GKE 1.37.2-gke.1042000 or higher, in `europe-west1` and `us-central1`.

Envoy is the foundation, not the product name. At Next '26 Google described Agent Gateway as understanding the MCP and A2A protocols, and said its agent networking work is ["built on an open foundation including Envoy and Kubernetes"](https://cloud.google.com/blog/products/networking/whats-new-in-cloud-networking-at-next26/).

Google and Tetrate engineers set out the technical case on April 3, 2026 in [Envoy: a future-ready foundation for agentic AI networking](https://cloud.google.com/blog/products/networking/the-case-for-envoy-networking-in-the-agentic-ai-era).

Envoy reads MCP and A2A message bodies, exposes tool and model names to RBAC and authorization filters, and can inject a user's delegated token so the agent never holds it.

Their example restricts an issue-triage agent to 3 read-only MCP tools on a server that exposes dozens. That is the gateway layer: a policy decision on every tool call, enforced at the wire.

## What is Chargebee Arc, and why does it matter for agent governance?

Chargebee [announced Arc on November 18, 2026](https://www.globenewswire.com/news-release/2026/11/18/3405953/0/en/chargebee-launches-arc-a-new-agentic-finance-platform.html) as "a new agentic platform that runs complete finance workflows as one connected, traceable thread." It opened for early access the same day.

Arc launched with 3 pre-built agents, for contract ingestion, receivables and revenue recognition, plus Agent Studio, a workspace for building custom agents on the same foundations.

The governance is in the commit step.

Per the release, "every action is shown, cited, and held for review before anything is committed," and human teams approve, edit or reject each proposal. [Arc's own site](https://witharc.ai) promises "a complete record of every decision: what was done, who did it."

That is the approval-gate layer: an agent may prepare a revenue schedule or a collections reply, but a person commits it, and the record says who. The release names no price and no general-availability date.

## What is Neo4j Relate, and what does it change about agent memory?

Neo4j [released Relate on November 11, 2026](https://neo4j.com/press-releases/announcing-relate-by-neo4j/) as "an AI-native, open source, in-memory and in-process embedded database," built in Rust and released under the Apache 2.0 license.

In Neo4j's words, "a single Cypher statement can search by vector similarity and execute graph traversals simultaneously." Reads are sub-millisecond in memory, with an optional fsync'd write-ahead log for durable stores.

Neo4j's [launch post](https://neo4j.com/blog/agentic-ai/meet-relate-by-neo4j-an-embedded-database-for-ai-agents/) says Relate runs as a library, a CLI or a WebAssembly build with no server, runs more than 30,000 test scenarios for openCypher, Cypher 25 and GQL, and can serve a database to agents as MCP tools.

That is the memory layer: agents store what they learn as connected facts rather than loose text, so the relationship between a customer, a contract and an exception survives the session.

Relate is the one launch of the three you can already own outright. It is open source and embedded, so it runs wherever your agents run.

## What do an agent gateway, an approval gate and an agent memory store have in common?

All three are the governance layer around an AI agent, and each decides something the model does not. The table sets the three November launches side by side.

<table style="width:100%; border-collapse:collapse; margin:1.5rem 0; font-size:0.95rem;">
  <thead>
    <tr style="background:#f5f5f0; border-bottom:2px solid #2175C5;">
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Layer</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">November 2026 launch</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">What it governs</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Where the rules and records live</th>
    </tr>
  </thead>
  <tbody>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Gateway</strong></td>
      <td style="padding:0.75rem;">Google Agent Gateway for cross-cloud (Limited Preview, Nov 11)</td>
      <td style="padding:0.75rem;">Which models, MCP tools and agents an agent may call, under which identity</td>
      <td style="padding:0.75rem;">Agents and MCP servers registered in Google's Agent Registry; free in preview, logging and Model Armor billed</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Approval gate</strong></td>
      <td style="padding:0.75rem;">Chargebee Arc (early access, Nov 18)</td>
      <td style="padding:0.75rem;">What a finance agent may commit: each action shown, cited and held for review</td>
      <td style="padding:0.75rem;">Chargebee's platform; no price or GA date published</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Memory</strong></td>
      <td style="padding:0.75rem;">Neo4j Relate (open source, Nov 11)</td>
      <td style="padding:0.75rem;">What an agent knows and how facts connect, in graph plus vector form</td>
      <td style="padding:0.75rem;">Wherever you embed it; Apache 2.0</td>
    </tr>
    <tr style="background:#f0f9ff; border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>All three, one stack</strong></td>
      <td style="padding:0.75rem;">ibl.ai: Agentic Router, Agentic OS, Agentic Ontology</td>
      <td style="padding:0.75rem;">Per-user tokens and roles, spend caps, human approval steps, per-user memory, certified tools, audit logs</td>
      <td style="padding:0.75rem;"><strong>Your infrastructure; you own all the code and the data</strong></td>
    </tr>
  </tbody>
</table>

The pattern matters more than any single product. The model is becoming the interchangeable part, and the controls around it are becoming the system of record for how your agents behave.

## Should an enterprise rent the agent governance layer or own it?

An enterprise should own the agent governance layer, because that layer holds the policies, the memory and the audit trail an examiner will ask for. Renting each piece separately puts those records in as many places as you have vendors.

Assemble the three November launches as sold and a single agent's rules sit in one vendor's registry, its approvals in a second vendor's application, and its memory in a database you embedded yourself.

Each is reasonable alone. Together they give you three policy models and three audit formats.

The gateway is the clearest case. Google's cross-cloud preview reaches into EKS, AKS and OpenShift, but the agents it governs are registered in Google's Agent Registry. Your agents can run on any cloud while their policy is held on one.

None of that makes the launches wrong. Identity per agent, deny-by-default tool access, a human approval before anything is committed, and memory as connected facts are exactly the controls to want. The question is only whose infrastructure holds them.

## How does ibl.ai run the agent gateway, memory and audit layers on infrastructure you own?

ibl.ai builds the three layers into one platform you deploy in your own cloud, on-premise or air-gapped, model-agnostic across any LLM. You own all the code and the data, under a perpetual license.

**The gateway.** [Agentic Router](/product/agentic-router) is one API key for 45 providers and 316 chat models (as of November 6, 2026) and 1,000+ MCP servers. Calls authenticate per user, with each person's own token and role, or server to server with a tenant-scoped key.

Spend caps apply at three scopes, the organization, one agent, and one person on one agent, and enforcement is server-side, so a runaway script stops at the cap. Cost is broken down per call.

**The memory.** [Agentic OS](/product/agentic-os) connects your systems of record over an MCP-based interoperability layer and assembles a secure per-user memory, with field-level permissions, PII redaction, RBAC, audit logs and prompt and response archives.

**The approval gate.** Agentic OS [workflows](/docs/os/chat-canvas/workflows) lay a multi-step agent run out as a graph, and a User approval node holds the run until a person signs off. A Guardrails node screens output before it moves on.

**The certified tools.** [Agentic Ontology](/product/agentic-ontology), open source under the MIT license, has your people certify the business terms agents call. A certified term becomes a tool with a fixed definition and fixed SQL, so an agent picks the right one and cannot widen it.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY, and our team stays on as a long-term partner, shipping updates and new agents after deployment.

For the same thesis applied to Broadcom's per-tool-call authorization, see [Agent Governance Moved Into Infrastructure](/blog/broadcom-agentminder-agent-governance-moves-to-infrastructure).

*Related: [AI Agent Governance: Managing Autonomous AI Systems Responsibly](/blog/ai-agent-governance-managing-autonomous-systems)*

## Why does owning the AI stack matter?

**ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.**

- **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours.
- **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform.
- **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.
- **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.

[Explore sovereigncorporation.ai](https://sovereigncorporation.ai/)
