Skip to content

Workflows

A workflow is a saved graph of steps that belongs to one agent. You ask the agent to create it in chat. Workflows exist only in General purpose workspaces.

In a graph, each step is a node and each arrow between steps is an edge.

Ask the Agent to Create the Workflow

The web app has no workflow editor. Send a chat message that asks for a workflow and describes the job. The agent follows a built-in skill before it saves anything.

  1. It runs the task by hand once.
  2. It moves reusable logic into a skill or a script when the logic is not trivial.
  3. It saves the graph.

If the workflow name already exists, the agent reports the conflict. It does not rename the workflow.

A Node Holds One Step

Each node is one step with a unique name. A node holds these fields.

Field What it holds
Goal What the step must achieve
Instructions What the agent does in the step
Done criteria How the agent knows the step is finished
Preferred skills Skills the agent loads before the step (optional)
Preferred tools Tools the agent favors in the step (optional)

Edges Connect the Nodes and Carry Branch Labels

An edge is one allowed move from one node to the next. The graph has no loops. It has at least one start node and one end node.

A node with several outgoing edges needs a branch label and a condition summary on every edge. The label names the outcome, for example reproduced or needs-info.

Inputs Are Typed Values or One JSON Payload

A workflow declares one input contract, either typed inputs or arbitrary JSON.

  • Typed inputs: a flat list of named values. Each value is a string, an integer, a number or a boolean.
  • Arbitrary JSON: one free-form JSON payload with a description and a default payload that prefills the form.

Schedules and webhooks check their values against this contract. A workflow with no parameters has no inputs.

The Graph Viewer Is Read-Only

  1. Select Workflows in the workspace sidebar.
  2. Pick an agent in Agent and a workflow in Workflow. The Graph tab shows the nodes and edges.
  3. Click a node. A panel shows its goal, Instructions, Done criteria, preferred skills and preferred tools.

The Workflows page with an agent picker and a workflow picker above a graph. A description card sits at the top left. Boxes are nodes and the lines between them carry labels such as no-duplicates and rubric-fallback

The description card names the workflow and shows its input count. Each box is a node, and each label on a line is a branch label.

You can drag nodes to tidy the layout. You cannot edit the graph. To change a workflow, ask the agent in chat to delete it and create it again.

The Model Follows the Graph

Each run starts a new session on the agent. The model of the agent reads the graph and carries out the nodes. For each node it reports Running, then Succeeded or Failed.

The agent does not ask you questions during a run. A workflow has no outputs field. The run keeps a status message, and the session keeps the transcript.

Warning

A run has no chat, so AgentZ allows tools marked consent required without a prompt. See Sandboxes.

Runs Move Through Five States and Obey a Time Limit

State Meaning
Pending The run exists and has not started.
Running The agent works through the nodes.
Succeeded The agent reported success.
Failed The agent reported a failure.
Unacked The session ended before the agent set a final status.

Each run has a timeout from 1 to 604,800 seconds (7 days). Schedules and webhooks default to 3,600 seconds. A run that exceeds its timeout records the reason TimedOut.

Next Step

Continue with Schedules.