Vertex AI Agent Engine and Dome
Deploy on Google Cloud, and govern every call the agent makes. Agent Engine, now Agent Runtime in Gemini Enterprise Agent Platform, hosts agent code you write. Give the agent a Dome key and point its tools and model at the Gateway. Dome decides every call and records it against the agent.
Dome governs every tool and model callin Vertex AI Agent Engine.
Your code, your endpoints
Agent Runtime runs the agent you deploy, in ADK, LangGraph or your own code. Its tools and model point at the Gateway.
The key in Secret Manager
The Dome key reaches the agent as an environment variable from Secret Manager. It is the only credential the agent holds.
Every call on the record
Tool and model calls land in one audit trail, against the agent. The same rules apply wherever else the agent runs.
Agent Engine on Domein three steps.
Register the agent in Dome, pass its key from Secret Manager, and point its tools and model at the Gateway. The Dome commands were run against a workspace; the Google Cloud side comes from Google's docs.
Register an agent and issue its key
Then store the key in Secret Manager.
shelldome agents register --name support-triager --gateway prod-gatewaydome agents create-key support-triager --name agent-runtimedome rules apply support-triager.cedar --agent support-triager --name linear-support-onlyPass the key at deploy
Agent Runtime's env_vars can reference a Secret Manager secret. The agent reads it from the environment.
pythonenv_vars = { "GATEWAY_URL": "https://<gateway-host>/gateways/<gateway-id>", "DOME_AGENT_KEY": {"secret": "dome-support-triager", "version": "1"},}Point tools and model at the Gateway
In ADK, use an MCP toolset for tools and LiteLLM for the model, with a pool as the model name.
pythonimport osfrom google.adk.models.lite_llm import LiteLlmfrom google.adk.tools.mcp_tool import McpToolset, StreamableHTTPConnectionParams GATEWAY_URL, DOME_AGENT_KEY = os.environ["GATEWAY_URL"], os.environ["DOME_AGENT_KEY"] tools = McpToolset(connection_params=StreamableHTTPConnectionParams( url=f"{GATEWAY_URL}/mcp", headers={"Authorization": f"Bearer {DOME_AGENT_KEY}"},))model = LiteLlm(model="openai/claude-opus-5-5", api_base=f"{GATEWAY_URL}/v1", api_key=DOME_AGENT_KEY)
File issuesfor Support only.
The permit lets the agent discover tools and call Linear. The forbid refuses any save_issue that doesn't name the Support team.
cedar
permit (principal, action == Dome::Action::"mcp:discover", resource);
permit (principal, action == Dome::Action::"mcp:call", resource is Dome::MCPTool)when { resource.connection_name == "linear" };
forbid (principal, action == Dome::Action::"mcp:call", resource is Dome::MCPTool)when { resource.connection_name == "linear" && resource.tool_name == "save_issue" && !(resource has arguments && resource.arguments has team && resource.arguments.team == "Support")};One call,two outcomes.
Switch the caller or the argument and watch the same call decide differently, with every decision landing in audit.
Team
agent support-triager · scheduled
linear/save_issue(team: "Support", title: "SSO login loops for Okta users")
- Agentsupport-triager is registered and active
- KeyBearer key from Secret Manager is valid
- RuleIssues for Support are allowed
DecisionAllowed
Bringing it alltogether.
Connecting agents on Vertex AI Agent Engine to registered tools, models, and identity in Dome completes a governed agent application.
Acting for
Agent
Vertex AI Agent EngineRuntime
Google ADKRuntimeControl point
GatewayEvery call decided and audited
RulesGuardsQuotas
Tools
Models
Common questionsabout Vertex AI Agent Engine.
More of whatDome works with.
Model
Claude Fable
Fable 5.1 from Anthropic and Amazon Bedrock in one failover pool, open to one group and capped by quota.
Provider
Anthropic
The Claude API behind the Model Broker: the key held in Dome, every call authorized, metered and audited.
MCP server
GitHub
The GitHub MCP server behind the Tool Gateway, with rules that decide each call on its owner and repository.
Identity
Okta
Okta tokens verified on every agent call, so rules and audit name the person each agent acted for.
Client
Claude Code
Claude Code on a Dome Gateway with per-developer sign-in, rules on every tool call, and audit by name.
Agent service
TinyFish
TinyFish's web agents behind the Tool Gateway, with rules that decide each run on the site it targets.
Talk with ourFDE team.
Or start free, and register your first agent in minutes. Connect it through a gateway, give it rules, and see every call it makes in one audit trail.
Talk with our FDE team
Our engineers design, build and run governed agents for the part of your business you want to change. You see what they deliver.
Agent frameworks and runtimes: govern agents without a rewrite
Build agents on the framework you already use. Dome governs their tool and model calls without a rewrite.