Workflow configuration

All step types, data mappings, branches, loops, joins, and the settings around the graph.

A workflow is an explicit graph. Each step performs work or makes a decision, and its outgoing connections determine what happens next. A valid graph also needs compatible inputs, outputs, and data mappings.

The workflow's settings and views

Area	What you manage
Workflow settings	Name, input contract, and output contract.
Start settings	Manual/API entry, webhook/event entry, schedules, and queue assignment for a published version.
First step	The node reached from Start.
Canvas and List	The same graph as a visual canvas or an ordered inspection of nodes and paths.
JSON	The full authored workflow definition, including settings not exposed by simple forms.
Data	Input, stored values, and step results available for mappings.
Runs	Earlier executions of this workflow.
Versions	Published snapshots; open a version or edit from it.
Triggers	Automation attached to published versions.
Client apps and Access	Application exposure and whether a subject can run the version and its dependencies.

Edits to a workflow draft autosave. Undo, Redo, and Discard draft affect authoring; they do not undo an external action already performed by a run. Publishing creates a new version.

Steps that do work

Step	Purpose	Main configuration
Agent	Run a published agent and use its answer.	Published agent, task mapping, result name, extra tools/skills, shared folder, and success/failure paths.
Workflow	Invoke a published workflow as part of this process.	Published workflow, input mapping, result name, and success/failure paths.
Tool	Call a registered read or write action.	Tool, arguments, result name, approval linkage where required, and failure path.
Send webhook	Send data to a reviewed destination using a tool step.	Destination, JSON body, approval where required, and failure path.
Set values	Put values into Workflow data for later steps.	Named values and the expressions used to compute them.

Send webhook is a convenience for a tool call, not a second delivery system. Destination setup and review still apply. Tool steps may also record that they undo an earlier tool step; this is an authored action, not automatic rollback of every external effect.

An Agent step can Also give it tools and Also give it skills for that invocation. Add a condition with Only when a rule holds, and choose whether a skill is always used or loaded when relevant. Whoever starts the workflow needs access to every listed resource, whichever way the conditions go. A tool requiring approval for each call belongs in a separate Tool step. These additions do not edit the agent's published definition.

Decisions and loops

Step	Purpose	Paths to connect
If	Test one condition.	Yes and No.
Cases	Choose the first condition that holds.	Each case and Otherwise.
Match	Compare one value against fixed choices.	Each choice and Otherwise.
Repeat	Repeat up to an explicit maximum.	Again and Rounds used up.
For each	Process each item of a list.	Each item and Finished; parallel mode also has join/failure behavior.

Simple comparisons include equals, not equal, ordering, membership, contains, and starts with. JSON expressions can combine conditions. Use the data picker to select the intended value and inspect its type before choosing the comparison.

A loop needs a deliberate path through its body and a terminating outcome. Make the limit-exhausted path useful rather than treating it as an unreachable edge.

Parallel work and joins

Fan out starts named branches concurrently. Each branch ends with End branch, which supplies its branch result. The join belongs to the fan-out configuration; it is not a separately authored work step.

Choose the join policy deliberately:

All must succeed: use when every branch's successful result is required.
Wait for all, report each: use when later steps need to inspect each outcome.
First to succeed wins: use when one successful answer is enough.

A For each can run one item at a time or run items in parallel with a join policy. Configure the item name, optional position name, saved joined result, and failure routing. Parallel branches do not automatically share every intermediate value; map the joined result for later work.

People, time, and endings

Step	Purpose	What to decide
Approval	Ask a person to decide on one proposal.	Subject, allowed reviewers, Approved/Rejected paths, optional Changes requested and timeout handling.
Wait	Pause for input or an external signal.	Wait kind, prompt or signal setup, response schema, stored response, and timeout/failure paths.
Timer	Pause for a duration.	Duration expression and the next step.
Output	Finish successfully with a value.	A result matching the workflow output schema.
Fail	Finish as failed with an explanation.	Failure code and message.
End branch	Finish one parallel branch.	The branch result; it does not replace the workflow's final Output.

A timer resumes an existing run. A schedule submits future runs. An external signal can resume a Wait, while an event trigger can start new work. Use the distinction to avoid accidentally starting duplicate processes.

Data mappings

Use Text or Fixed for a literal value, From data for input or earlier results, and JSON for a more complex expression. For text, Insert a value helps combine text with a selected value.

Data source	Example use
Workflow input	Pass the submitted request to the first agent.
Saved step result	Pass the agent's answer into an approval proposal.
Workflow data	Reuse a value created by Set values.
Current loop item	Process the document in this For each iteration.
Joined branch result	Combine completed parallel work.

The source must exist on the path that reaches the consuming step. A branch-specific value cannot be assumed present on every path. Changing an agent's output shape can require new mappings and a new workflow publication.

Before publication

Check Start, every required outgoing path, reachable ending steps, loop maxima, published agent/workflow selections, tool access, approval links, and schema compatibility. Use the reported problems to find the node to repair. Then run representative success, rejection, timeout, and failure paths before exposing the version to a client app or automation.