Workflows

Workflows let you chain agents and tools into repeatable, multi-step pipelines — all designed in a visual canvas inside the app. Where a single agent handles one conversation turn, a workflow can gather data from multiple sources, pass it through several agents, apply conditional logic, and deliver a structured result from a single trigger.

Opening the workflow builder

Go to Workflows in the left sidebar and click New workflow. The visual workflow canvas opens with an empty starting step ready to configure.

Steps

A workflow is made up of steps. Each step appears as a card on the canvas. Drag cards to reorder them; connect them by drawing arrows between the output port of one card and the input port of the next. Click a step card to open its configuration panel on the right.

Agent step

Invokes one of your saved agents with a message you compose in the step configuration panel. The agent runs its full reasoning and tool-call loop and returns a text output that subsequent steps can reference. Use agent steps for any task that requires reasoning, synthesis, or language generation.

Tool step

Calls a single built-in or dynamic tool directly — without involving a language model. Use tool steps for deterministic actions where no reasoning is needed, such as fetching a URL, running a calculation, or querying an API.

Transform step

Reshapes data between steps using a template. In the configuration panel you write a template that can reference outputs from earlier steps, extract fields from a JSON response, concatenate strings, or format a final result for display. No AI call is made, so transform steps run instantly and at no cost.

Branch step

Evaluates a condition and routes execution to one of two paths — Yes or No. In the configuration panel, use the condition builder to compare an output value from an earlier step against a threshold or label (for example, checking whether a confidence score is above 0.8 or whether a classification returned "urgent"). Each branch connects to a different downstream step on the canvas.

Passing data between steps

Every step produces named outputs. In any downstream step's configuration panel you can reference those outputs by selecting them from the Insert variable dropdown. The builder inserts the correct reference for you, so you don't need to remember the exact syntax. The output of one agent step can become the input message for the next, or a field extracted by a transform can be passed as a parameter to a tool step.

Workflow inputs

Define the information a workflow expects at run time using the Inputs panel at the top of the canvas. Add each input with a name, type, and optional default value. Inputs appear in the trigger form when you or a teammate runs the workflow, and they are available as variables throughout all steps.

Running a workflow

To run a workflow, open it in the sidebar and click Run. A form appears with fields for any inputs you defined. Fill them in and click Start.

The execution panel shows each step completing in real time — green for success, red for error. Click any step card to see its inputs, outputs, and elapsed time. If a step fails, the error message appears inline and the rest of the workflow halts.

You can also trigger a workflow from inside a chat by typing /run [workflow name] in the message input. The workflow result streams back into the conversation.

Run history

Every execution is saved in the workflow's History tab. Select any past run to replay it with the same inputs, inspect the full step trace, or download the output. Run history is useful for auditing results and debugging flaky steps.

Reusing agents and tools

Any agent or tool defined in your workspace can be referenced in any number of workflows. If you update an agent's system prompt, every workflow that uses that agent picks up the change immediately. For this reason, it is good practice to test significant agent changes on a duplicate before updating the original — this avoids disrupting workflows that are already in use.

Best practices

• Name steps clearly — the step name appears in output references and error messages. Descriptive names like "Summarise article" or "Quality gate" are far easier to debug than "Step 3".
• Use transform steps to shape data — avoid asking an agent to reformat data when a transform step can do it instantly and at no cost.
• Test both branch paths — before relying on a workflow, run it with inputs that exercise both the Yes and No sides of every branch to confirm both paths behave correctly.
• Use focused agents in workflows — a specialised agent with a narrow system prompt performs more reliably as a workflow step than a general-purpose agent asked to do everything. See the Agents guide.
• Schedule recurring workflows — in the workflow settings, configure a schedule to run the workflow automatically at a fixed interval — useful for regular reports, data refreshes, or monitoring tasks.