Agents overview

What an agent is in VeloPhex, how a run works, what you need to turn Agents on, and who can do what.

BetaAgents is in beta. It needs two things: The agentic feature switched on for the Orchestrator. It is off by default: set the configuration key Features:Agentic to true (as an environment variable, Features__Agentic=true) and restart the Orchestrator. The vector store needs the pgvector extension in the Orchestrator's PostgreSQL database. The Agents service enabled for the tenant by a platform administrator, and included in the tenant's subscription. See Platform administration. While the feature is off, the agentic pages are not offered and their API answers 404. While the service is off for the tenant, the agent API answers 403; runs that are already going keep working. An agent is a model that calls tools until it has an answer. You write its instructions, pick its model and choose the tools it may use: your processes and your vector stores. VeloPhex runs the agent as a job on one of your Robots, records every step, and stops before any call that needs a person's approval. Open Agents from the app launcher in the top-left corner of the Orchestrator, or go to /agents. Agents opens in your current workspace. Concepts# Term Meaning Agent A named definition in a workspace: Instructions, Model, optional Input schema and Output schema, and Tools. Each agent has a Key (for example invoice-triage-agent) used in URLs and the API. Draft What you edit. Every save is a new revision. Version A published, unchangeable snapshot of a draft: the definition plus exactly what each tool and the model resolved to when it was published. Runs started from outside use the newest published version unless they name another. Run One execution of an agent. A run is a job of the workspace's Agent Runner process, so it also appears under Automations ▸ Jobs. Kinds: Test (from the builder), Invoke (from the API) and Evaluation. Step One thing the agent did in a run: a model call, a vector store search, a tool call or an approval. Approval A tool call waiting for a person to approve or reject it, with its exact arguments. Model A named model binding such as agent:default, pointing to an AI connection and a provider model. AI connection The tenant's credentials for a model provider. Keys stay in the Orchestrator; Robots never see them. How a run works# Someone starts a run: Test agent in the builder, or a call to the API. The run waits as a job until a Robot with the Python runtime claims it. The Robot runs the workspace's Agent Runner process. The agent calls its model through the Orchestrator, which uses the tenant's AI connection. When the model asks for a tool, the agent calls it. A search of a vector store runs at once. A process tool waits for a person to select Approve call. The agent returns its answer. If the agent has an output schema, the answer is checked against it. Every step is shown on the run page as it happens, with tokens, duration and cost. What you need# Requirement Where Page Features:Agentic on, and the Agents service enabled for the tenant Orchestrator configuration, Platform administration Set up Agents An AI connection Orchestrator Settings ▸ AI connections Set up Agents A model of purpose agent Agents ▸ Models Set up Agents A Robot with the Python runtime in the workspace Orchestrator Robots Machines and robots The Agent Runner process in the workspace Orchestrator Automations ▸ Processes Set up Agents Who can do what# Agent permissions are workspace permissions. Add them to a role. See Manage access. Permission Allows tools.view Open agents, versions, models and approvals. tools.create Create agents. tools.edit Edit, test and publish agents; approve or reject tool calls; add or override models. tools.delete Delete agents. jobs.view See runs. jobs.create Start test runs and invoke agents. jobs.cancel Stop a run. traces.view Read the steps of a run and execution traces. evaluations.view, evaluations.create Read evaluations; start an evaluation of an agent version. vector-store.view, vector-store.create, vector-store.delete Use, create and delete vector stores. mcp_servers.view, mcp_servers.create, mcp_servers.edit, mcp_servers.delete Manage MCP servers. credential_stores.view, credential_stores.edit, credential_stores.delete Read, add and change, and delete AI connections (tenant-wide). The Connections link in the Agents menu is shown only to people with credential_stores.view. Next steps# Set up Agents Build an agent Test and publish an agent

Concepts

How a run works

What you need

Who can do what

Next steps