Google Gemini Enterprise Agent Platform vs Orchestrator Zero
oz0 records agent jobs running on your infrastructure. Agent Platform hosts agents on Google Cloud alongside Gemini and Model Garden. oz0 provides a control plane you host, outbound-only nodes on your machines and access to local and cloud models.
What Google Gemini Enterprise Agent Platform is
Gemini Enterprise Agent Platform is Google Cloud's platform for building, deploying, governing and optimising agents, and the successor to Vertex AI and its Agent Engine. Agents built with ADK or other frameworks run in Agent Runtime, a managed container runtime, with Sessions, Memory Bank and code execution sandboxes, while Agent Gateway, Agent Identity and Agent Registry govern which tools and agents they can reach. Models come from Model Garden, which holds Gemini and third-party and open models.
docs.cloud.google.com- Runs
- Managed service
- Checked
- 27 October 2026
Side by side.
Google Gemini Enterprise Agent Platform's ratings come from its own pages. A column's name opens its definition and how Orchestrator Zero does it. How we rated
Where it runs
Any machine you own
Yes
DocsConditional: Google-managed runtime
Agent Runtime is a managed runtime that Google operates; you do not choose the machines agents run on.
Linux, macOS and Windows
No: Google cloud only
Agents run in Agent Runtime as containers that Google manages; nothing runs natively on your own Linux, macOS or Windows machines.
Self-hosted control plane
Yes
DocsNo: Managed Google Cloud service
Agent Runtime is a managed Google Cloud service; there is no control plane to run yourself.
No inbound ports
Yes
DocsNo: Container serves port 8080
The agent container must serve HTTP on port 8080, and the platform calls into it.
Fleet operations
Yes
DocsNo: Fully managed runtime
The runtime hides the infrastructure; there are no machines of yours to join, label, route to or block.
Durability
Durable runs
Conditional: Manual resume by invocation ID
ADK can resume a stopped workflow, but someone has to resume it with the invocation ID; the platform does not continue it by itself.
Governance
Tool policy outside the model
Conditional: Allow and deny, no approve
Admins write IAM Access policies that Agent Gateway enforces on each MCP call, with allow and deny rules on tool names and annotations and deny by default, but no rule can make a call wait for a person; plain-language semantic governance policies are a preview.
docs.cloud.google.com docs.cloud.google.com docs.cloud.google.com
Approvals in chat
Conditional: Experimental, in ADK code
ADK's tool confirmation pauses a tool for a yes or no from the ADK web UI, CLI or REST API, but it is experimental and does not work with the platform's managed sessions; no chat app approvals are documented.
Tool snapshots
Yes
DocsConditional: Registry copy; manual updates
Agent Registry stores uploaded MCP tool definitions. The docs describe manual registration and updates, not an automatic comparison that blocks changed tools pending approval.
Tenants built in
Yes
DocsConditional: IAM per user, no tenants
IAM conditions can limit each user to their own sessions, but there is no tenant construct.
Cost
Budgets that stop a run
Conditional: Token metrics, no budget
Token usage is reported as a metric, but no budget stops a run.
Security
Provider keys off the workers
Yes
DocsConditional: IAM identity, no static key
Agents call Gemini and Model Garden models with their agent identity or service account, which the runtime holds; API keys for other services come from the auth manager, a preview, and ADK adds them to requests inside the agent process.
Sandbox with an egress allowlist
Conditional: Allowlist only in preview
Code Execution sandboxes are isolated and have no network access; the Managed Agents API sandbox takes a domain allowlist but is a pre-GA preview, and other egress control goes through your own VPC.
docs.cloud.google.com docs.cloud.google.com docs.cloud.google.com
Audit and compliance
Tamper-evident audit trail
Conditional: Audit logs, no integrity proof
Cloud Audit Logs record each Agent Gateway authorization decision and the entries are immutable, but Google documents no hash chain or signature you can verify.
Erasure by key
Conditional: Purge per user, no own keys
Memory Bank can purge every memory in one user's scope and sessions can be deleted, but no encryption key per person is documented.
Retention with legal hold
Conditional: TTL settings, no legal hold
Sessions and memories take a time to live (365 days by default for sessions), but no legal hold is documented.
Openness
Model provider choice
Yes
DocsYes
Model Garden offers Gemini and over 200 foundation models from Google, partners and open-source projects, and ADK agents can also use models hosted elsewhere or running locally.
MCP and A2A
Yes
DocsYes: A2A hosting in preview
Agent Gateway carries MCP and A2A traffic and registered A2A agents can be used as subagents; hosting an A2A agent on Agent Runtime is a preview.
Where Google Gemini Enterprise Agent Platform is stronger.
A visual agent canvas
Agent Studio is a low-code canvas for designing agents and their workflows.
docs.cloud.google.comOver 200 models in one catalog
Model Garden gives one catalog of Gemini, partner and open models, with managed serving and tuning in the same project.
docs.cloud.google.comManaged runtime, fast starts
Agent Runtime runs agents in any language as managed containers with sub-second cold starts and long-running sessions.
docs.cloud.google.comCertificate-bound agent identity
Each agent gets a SPIFFE identity whose tokens are bound to mTLS certificates and DPoP, so a stolen token cannot be used from another machine.
docs.cloud.google.com
Differences to examine.
Compare the scope, conditions and editions in the source notes.
- Any machine you own
- Linux, macOS and Windows
- Self-hosted control plane
- No inbound ports
- Fleet operations
- Tool snapshots
- Tenants built in
- Budgets that stop a run
- Provider keys off the workers
- Tamper-evident audit trail
- Erasure by key
Using both
An agent on Agent Runtime can reach an Orchestrator Zero agent through Agent Gateway as an MCP server or an A2A agent, since every Orchestrator Zero agent is both, under the gateway's access policies.
Based on documentation; this pairing has not been tested.
Find your first task.
Explore an example or read the public Docs preview. Software access is coming soon.