Production agent workflow
KeyRunner does not replace your agent framework. Your application continues to own prompts, reasoning, and user experience. KeyRunner provides the governed tool catalog and execution path used by that agent.
Prepare the agent
- Export tested requests as tools.
- Group the required tools in one or more policies.
- Create or select an agent in AI Workspace.
- Assign policies and review the effective tool count.
- Configure the supported authentication or workload identity.
Connect from code
Use the integration details on the agent screen to connect through the KeyRunner SDK or MCP endpoint. SDK workflows initialize the KeyRunner client and retrieve the tools available to the current agent. MCP-compatible clients use the provided endpoint and authentication configuration.
Do not embed production credentials in prompts or client-visible configuration. Keep provider keys and API credentials in environment variables or supported secret stores.
Operate the agent
The agent overview summarizes success rate, average latency, production tools, and recent activity. Activity is searchable so developers can find a trace, tool, policy, status, or error without leaving the agent context.
Before removing a policy, review the impact dialog. A tool becomes unavailable only when no other assigned policy grants it.
Diagnose production behavior
Open Observe for workspace-wide traces or select an agent activity entry for details. Use the trace graph when an execution includes multiple model or tool calls. See Agent observability and trace graph.