Google Thinks Agents Need Their Own Enterprise Stack

Google’s launch of Gemini Enterprise Agent Platform is less interesting as product news than as market evidence. The company is making a fairly blunt argument: agents are no longer something you bolt on to a model, or tuck inside a chatbot, or ship as a thin SDK. They now require their own enterprise stack.

That is the significance of the announcement. Google is not merely adding another agent framework to an already crowded pile. It is trying to define the control plane on which enterprise agents are built, run, supervised and measured.

The package is expansive. There is an updated Agent Development Kit, a visual Agent Studio, long-running runtime support, persistent memory, sandboxed execution, agent-to-agent orchestration, simulation, evaluation, observability, and a governance layer built around Agent Identity, Agent Registry and Agent Gateway. Read together, those pieces amount to a claim that the awkward parts of agent adoption, the ones that begin after the demo, should now be treated as platform concerns.

This is not just another framework launch

The easiest mistake is to read this as another model company dressing up workflow tooling in more ambitious language. Google is doing something more deliberate than that.

The platform combines four layers that are often spread across different products and teams. There is agent construction: code-first and low-code tooling, templates, integrations and model choice. There is runtime: long-running execution, sessions, memory and sandboxed environments for code and browser work. There is governance: identity, access control, gateways and policy enforcement. And there is optimisation: simulation, evaluation, observability and execution traces.

That matters because the hard part of enterprise agents is rarely getting the first one to work. The hard part is deciding who can approve tool access, how long an agent can run, what it is allowed to remember, how its actions are traced, how it is tested, and what happens when it behaves strangely at 2am.

The comparison that matters: platform versus product versus runtime pattern

Comparing Gemini Enterprise Agent Platform with OpenClaw, NemoClaw, Hermes and Microsoft 365 Agents is slightly untidy because these systems are not all trying to solve the same problem.

Google is building a managed enterprise platform. Microsoft 365 Agents are an extension model inside a productivity suite. OpenClaw and Hermes are closer to self-hosted agent operating environments. NemoClaw is a hardening and deployment layer built around OpenClaw.

Still, the overlap is substantial enough to make the comparison useful. All of them are, in different ways, answers to the same underlying question: what does it take to move from an AI assistant that can answer to an AI system that can act?

High-level comparison

Stack	Best understood as	What it does well	What it does not really try to do
Gemini Enterprise Agent Platform	Managed enterprise agent platform and control plane	Multi-model agent building, long-running runtime, memory, sandboxing, governance, observability, evaluation, enterprise integrations	Local-first personal agent computing, lightweight hobbyist setup, deep user-owned self-hosting by default
OpenClaw	Self-hosted agent runtime for personal and team workflows	Messaging-native agent, typed tools, cron, browser and device control, memory, subagents, automation in your own environment	Out-of-the-box enterprise governance plane, central registry, formal fleet identity, turnkey security dashboards
NemoClaw	Hardened deployment stack for OpenClaw	Sandboxed execution, routed inference, declarative network policy, safer unattended OpenClaw deployment	Broad first-party agent feature invention beyond OpenClaw itself, low-code business-user builder UX
Hermes	Broad self-hosted agent platform with strong user features	Rich tool system, skills, memory, browser automation, voice, batch runs, MCP, provider routing, ACP and API server support	Enterprise-native control plane with central governance for a whole regulated organisation
Microsoft 365 Agents	Agent extension model inside Microsoft 365 Copilot and Copilot Studio	Tight fit with Microsoft 365 apps, enterprise data, connectors, productivity workflows and admin controls	General-purpose self-hosted agent runtime, deep local machine autonomy, model-agnostic platform independence

Capability snapshot

Stack	Governance and security	Memory model	Orchestration style	Deployment model
Gemini Enterprise Agent Platform	Strong first-party governance story with identity, registry, gateway, policy controls and observability built in	Persistent platform memory via Memory Bank and sessions	Multi-agent orchestration with both generative and deterministic patterns	Fully managed Google Cloud platform
OpenClaw	Practical guardrails and approvals, but governance is shaped by your config rather than a central enterprise control plane	File-backed and session-aware memory designed for continuous assistant workflows	Subagents, cron, tool use and workflow routing inside a self-hosted runtime	Self-hosted in your own environment
NemoClaw	Stronger operational security than plain self-hosting through sandboxing, routed inference and declarative egress policy	Inherits OpenClaw memory behaviour while adding safer runtime boundaries	Mostly follows OpenClaw orchestration patterns, with emphasis on safe unattended execution	Self-hosted, but wrapped in NVIDIA’s hardened deployment stack
Hermes	Good operator-level controls, plugins and provider routing, but less opinionated about central enterprise governance	Persistent memory plus optional external memory providers	Delegation, batch processing, hooks and broad tool orchestration	Self-hosted and builder-oriented
Microsoft 365 Agents	Strong admin and compliance fit inside the Microsoft estate, especially where the Microsoft Graph is already the centre of gravity	Grounded in Microsoft 365 data, connectors and enterprise context more than an open-ended memory subsystem	Workflow and agent orchestration inside Copilot Studio and Microsoft 365 surfaces	Managed inside Microsoft 365 and Copilot Studio

Google is aiming at a different layer from most open systems

Open systems such as OpenClaw and Hermes are compelling because they make agents feel operational now. You can wire them into chat, tools, local files, browsers, cron jobs and custom workflows without waiting for a central platform team to bless every move. That is why they matter. They close the gap between AI that talks and AI that actually does work.

Google is aiming at a different buyer. The buyer here is the organisation that wants a standardised control plane for many agents across many teams, with auditability, security policy, evaluation and fleet-wide governance treated as product features rather than side projects.

In that framing, OpenClaw and Hermes look less like direct competitors than evidence that the category is maturing. They have already shown what persistent agents, multi-step automation, memory, tools and delegation can look like in practice. Google is packaging the enterprise version of that idea.

OpenClaw still has an advantage Google does not

Where OpenClaw remains stronger is proximity to the user and the environment. It is built for agents that can live inside an actual workflow: Discord, Telegram, Slack, cron, browser control, local files, device integrations, subagents and task routing. It feels like an agent runtime you shape around your work, not simply a managed cloud product you configure through a console.

That distinction matters. Some organisations want a governed estate of enterprise agents. Others want an assistant that can inhabit their stack, run where they need it to run, and remain legible because they control the whole system.

Google’s platform is stronger if the problem is standardisation at scale. OpenClaw is stronger if the problem is usefulness, adaptability and operator control in a user-owned environment.

NemoClaw is the bridge from interesting to acceptable

NemoClaw deserves separate attention because it is not just OpenClaw with a different badge. It is a hardening layer designed to make OpenClaw-style agents easier to run in unattended or enterprise-sensitive settings.

That means sandboxing, routed inference, egress control, lifecycle management and policy-based guardrails. In practice, NemoClaw is tackling the objections infrastructure and security teams tend to raise the moment they see an autonomous agent with shell access and broad network reach.

So while Google is offering a top-down managed platform, NemoClaw is trying to make a bottom-up open runtime easier to defend. It does not remove the need for enterprise operations, but it narrows the gap between experimental agent systems and something a serious operator might accept.

Hermes is closest in breadth, but not in buyer

Hermes is probably the closest match in raw capability breadth from the self-hosted side. Its toolsets, skills, memory, browser automation, voice support, batch processing, MCP integration, provider routing, ACP support and API server make it feel like a full agent environment rather than a single-purpose assistant.

But Hermes still reads as a builder and power-user stack before it reads as an enterprise control plane. That is not a flaw. It is a different optimisation target. Hermes is designed to be expressive, hackable and broad. Google is optimising for standardisation, governance and managed rollout across large organisations.

Microsoft 365 Agents are narrower, but strategically so

Microsoft’s answer is narrower, but strategically sensible. Microsoft 365 Agents live close to the Office graph, Teams, Outlook, documents, connectors and the admin fabric enterprises already pay for. That gives Microsoft a distribution advantage Google cannot casually copy.

The trade-off is that Microsoft 365 Agents are less about becoming a universal agent runtime and more about extending a specific enterprise work surface. If a company already lives inside Microsoft 365, that may be enough. If it wants broader cross-system autonomy, model flexibility and runtime-level control, the shape is less obviously sufficient.

What builders should take from this

The market is beginning to split into three camps.

One wants managed enterprise agent infrastructure. That is where Google is placing its bet.

Another wants agents embedded inside existing productivity suites, with familiar connectors, admin tooling and procurement paths. That is Microsoft’s lane.

The third wants open, self-hosted, operator-shaped systems that can be adapted quickly and run close to the work. That is where OpenClaw, NemoClaw and Hermes remain most interesting.

The likely outcome is not a single winner. It is a layered market.

Large companies will buy control planes. Technical teams will continue to experiment with open runtimes. The interesting middle ground will be occupied by products that can combine the freedom of open agent systems with enough governance to satisfy a real organisation.

That is why Google’s launch matters. It confirms that agents are moving beyond model wrappers, chatbots and workflow demos. They are becoming infrastructure. The only question is whose infrastructure organisations decide to trust.

Sources

• Google Cloud: https://cloud.google.com/blog/products/ai-machine-learning/introducing-gemini-enterprise-agent-platform
• NVIDIA NemoClaw docs: https://docs.nvidia.com/nemoclaw/latest/about/overview.html
• Hermes docs: https://hermes-agent.nousresearch.com/docs/user-guide/features/overview/
• Microsoft Learn, agents for Microsoft 365 Copilot overview: https://learn.microsoft.com/en-us/microsoft-365-copilot/extensibility/agents-overview