Skip to main content

Workflows: build

For Studio users building a repeatable process on the board. You need an active agent for every Task (Agents); an AMPG update step needs a graph version with training documents; an S3 prefix parameter needs an AWS connector on one of the workflow's agents (AWS connector).

Steps​

Open the builder​

  1. In Workflows, click Create Workflow. To change an existing workflow, open it and click Create Version in Workflow Versions; the builder opens with the current graph.
  2. Fill Name this workflow (80 characters) and What it does, in one line (1,000).

Add and configure steps​

  1. Click a type in the Add a step rail, or drag it onto the canvas (empty: "Add a Task to start."). The five types are on the Workflows hub; each card's footer reads Asks you or Runs unattended. The model is set by the type: the drawer reads Model with the id the step runs on and "set by the step type", the Coordinator model for a Conditional, AMPG update or Redact PII step and the Worker model for a Task (for example anthropic.claude-sonnet-5-5); a Handoff runs no model and shows none; if the ids cannot be read, it shows the model's class and the reason.
  2. Click a card to open Configure followed by its handle. Every step has a Title; a Handle ("How other steps refer to this one": lowercase letters, digits and dashes, unique; renaming rewrites every mention in other steps, connections, branch labels and AMPG sources); a Things to fix on this step list while incomplete; and Delete step. Escape closes the drawer.
  • Task: pick the agent (Find an agent; rows show its knowledge graph and environment); write the Objective ("What should the agent do here?"), typing @ to refer to a connected earlier step's result; describe the output ("What comes out - a memo, a table, a chart…") and its kind: text, pdf, docx, pptx, xlsx, data or any. Two toggles make it ask you: Hold the result for my review (an Inbox card shows what it produced with Approve · Send back · Stop) and Wait for my go-ahead before it runs (an Inbox card shows its objective with Approve & run · Send back · Stop).
  • Handoff: write the message you will read ("What should you check before the run continues?"). Who acts · You (the workflow's owner); outcomes: Approve & run continues past the step; Send back gives your note to the step before, which runs again; Stop cancels the run here and everything produced so far stays in Files.
  • Conditional: write what the decision is based on ("An evaluator on anthropic.claude-opus-5-5 reads the upstream results and picks exactly one branch.", naming your deployment's Coordinator model). Each outgoing connection is a branch, labelled under Branches ("Label for the branch to @x") and shown as a pill on the connection; at least two with distinct labels.
  • AMPG update: Knowledge-graph update node. Redact PII: PII redaction.

Wire steps​

  1. Drag from a card's right handle (Connect to a later step) to another's left handle (Connect from an earlier step); hover a connection and click Remove connection to delete it. A connection from a step to itself, or one that would make the workflow loop back, is not made. Delete or Backspace removes a focused card; the Delete step confirmation names it and removes its connections, branch labels and AMPG sources.

Settings and parameters​

  1. Click Settings for the Workflow settings popover (click outside or Escape to close):

    • Parallelism: steps running at once, 1 to 10; default 3.

    • If a run is already going: Skip, Queue or Parallel, applied to a schedule or a second Run Now. With Queue, new runs wait and start one at a time, oldest first, once the run before them finishes; at most 10 can wait, and beyond that a further Run Now is refused and a further schedule firing is skipped.

    • If a step fails: Stop on first failure or Continue where possible.

    • Time limit: off by default; switching it on sets 4 hours; type 0.25 to 168 or pick 1 h, 4 h, 24 h or 72 h. "The run is stopped when it has been running for this long, including time spent waiting for your approvals. Off means no limit."

    • PII redaction: PII redaction.

    • Parameters, available to steps as {{params.name}}. Add parameter, then per parameter: a name (a lowercase letter, then lowercase letters, digits or underscores, up to 64), a type, and whether it is required, its default and a description ("What the caller should pass"). Empty: "No parameters - Run Now starts the workflow straight away."

      TypeExtra fields
      TextOptional minimum and maximum length; optional whole-value pattern
      NumberOptional minimum and maximum
      Yes / noNone
      ChoiceThe values, comma-separated
      S3 prefixThe AWS connector whose role reads the bucket (only connectors on the workflow's agents); optional Fixed bucket and Default prefix; Max objects (default 500, at most 5,000) and Max bytes (default 2 GiB, at most 20 GiB); the objects are copied into the run before the first step

    • Result schema (JSON Schema), up to 32 KB, checks outputs/result.json.

Save, activate, discard, delete​

  1. The top-bar chip reads Ready to save or N things to fix (hover for each). Save as version creates a new version and returns to the workflow's page; a server refusal shows word for word in a red bar.
  2. In Workflow Versions (Version, Notes, Created, Status; "Each save creates a new version. Activate a version to make it the one used by Run Now and any active schedules.") click Activate; Active Version shows it and the badge turns Active.
  1. Cancel with unsaved changes asks Discard changes ("Your unsaved changes will be lost.", Discard); closing the tab raises the browser's warning. Delete on the workflow's page opens Delete Workflow: type the name, then Delete. Studio then shows an empty builder; choose Workflows in the sidebar to return to the list.

What you should see​

After a save: a new row in Workflow Versions; Active Version reads None and the badge Inactive until you activate. After Activate: Status Active; Run Now, Run on Schedule and Share available; Workflow Preview shows the active graph read-only. Other badge values: Draft · ready to activate on a Library copy with its dependencies wired; Needs attention when a pulled dependency could not be resolved (hover for the reason).

Notes​

  • Version History (/workflows/<workflow id>/versions) offers Edit Workflow, View Runs and Deactivate (clears the active version), shows each version's Health (Healthy, Degraded or Broken; hover for the issues), and greys out Activate with the reason. Steps do not pin an agent version: each run uses the agent's active configuration at the time it runs, so a change to an agent applies from its next run. A version whose agent lost something it needs, such as its execution environment, reads Broken and cannot be activated.
  • On a copy pulled from a Library, Deactivate and then Activate in Version History shows the version active there, while the workflow's page and the list keep reading Draft and Inactive. Run Now still starts runs.
  • Limits: 20 steps ("This workflow has the maximum of 20 steps."); 256 workflows per user ("You've reached the limit of 256 workflows." disables Create Workflow; a governed deployment lists everyone's, so the server enforces your own count); one agent per Task; two labelled branches per Conditional; one source per AMPG update. An agent, connector or knowledge graph used by an active version cannot be deleted or repointed: This primitive is in use lists the workflows, by workflow ID, and their versions ("Open each affected workflow and detach this primitive (or activate a different version) before retrying."); a draft workflow that names it is not listed.
  • Fix messages name the step ("Choose the agent that does this step.", "Say what the agent should do.", "Give the step a title.", "Label the branch to @x.", "The workflow loops back on itself through @a, @b."). A Task reading "Not in your active agents": pick another agent or reactivate that one. "Could not load the workflow.": reload; every URL survives a reload. "Cannot activate: …": hover Health, fix in a new version.