Set up Agents

Turn the agentic feature on, connect a model provider, check the models, and register the Agent Runner so a workspace can run agents.

Set up Agents once per deployment, once per tenant and once per workspace. The Agents page shows a checklist until the workspace is ready. The checklist names the first step that is missing and links to where you fix it: Add connection, Add model, Open robots or Open processes. It disappears when all four are done. Turn on the agentic feature# Agents and the shared AI features (AI connections, models, vector stores, MCP servers, evaluations and traces) are part of the Orchestrator but switched off by default. Make sure the Orchestrator's PostgreSQL database has the pgvector extension available. The vector store tables use it. Set Features:Agentic to true in the Orchestrator configuration, for example as the environment variable Features__Agentic=true. Restart the Orchestrator. The workspace menu now has an Agentic group with Vector stores, MCP servers and Evaluations. Enable the Agents service for the tenant# A platform administrator enables the service. See Platform administration. Open Platform administration ▸ Tenants and select the tenant. Open Services and select Enable service for Agents. Check that the tenant's subscription includes Agents. Add an AI connection# An AI connection holds the credentials for one model provider. It belongs to the tenant, and the key is write-only: after you save it, the web app only shows Key set. You need credential_stores.edit (for example the tenant Administrator role). In the Orchestrator, open Settings ▸ AI connections. From Agents you can also select Connections in the menu. Select Add connection. Enter a Name. Choose the Provider: Provider Also asks for OpenAI API key Azure OpenAI Endpoint, API version, API key Anthropic API key OpenAI-compatible Base URL; API key is optional. Switch on Private endpoint when the endpoint is inside your network. Keep Enabled on and select Save connection. On the connection's ⋮ menu, select Test connection. The status changes to Connected, or shows why the provider refused the call. The provider of a saved connection cannot be changed. To replace a key, open ⋮ ▸ Edit, select Replace next to Key set, and enter the new key; leaving the field blank keeps the current key. The first connection you add also creates three tenant-wide models that use it: agent:default, embedding:default and judge:default. WarningDeleting a connection stops every model that uses it. Check the models# A model binding is addressed as purpose:name, for example agent:default. Agents use bindings of purpose agent; vector stores use embedding; evaluations that ask a model to judge use judge. In Agents, select Models. Check that there is a binding of purpose agent (normally agent:default). Scope shows Inherited for the tenant's models and Workspace override for this workspace's own. To use a different provider model in this workspace: On the binding's ⋮ menu, select Override (or Edit for an existing override). Choose the Connection and enter the provider's Model (or Azure deployment) name. Select Save model. To add a new binding, select Add model, choose the Purpose (agent, embedding or judge), enter a Name (lower-case letters, digits, dots, hyphens or underscores), a Connection and a Model, and select Save model. ⋮ ▸ Remove override returns the workspace to the tenant's binding. Changing models on this page needs tools.edit in the workspace. Check the Robot# An agent run needs a Robot with the Python runtime, on a machine assigned to the workspace, and a robot account in the workspace. See Machines and robots. Register the Agent Runner# Agent runs are jobs of a process built from the Agent.Runner package. Each workspace that runs agents needs exactly one such process. Upload the Agent.Runner package from your VeloPhex release: Orchestrator Packages ▸ Upload. See Packages and processes. In the workspace, open Automations ▸ Processes and create a process from the Agent.Runner package with the entry point Run agent (agent). Name it, for example, Agent Runner. The checklist step Runner registered is now done. If the workspace has no Agent.Runner process, or more than one, the checklist says what to fix. When you install a new Agent.Runner version, upload it and point the process at it. Versions that were tested on the old runner must be tested and published again. Next steps# Build an agent Create a vector store

Turn on the agentic feature

Enable the Agents service for the tenant

Add an AI connection

Check the models

Check the Robot

Next steps