Workflows.

Start a workflow

• A signed-in account. Workflows is in the rail.
• A workspace open, to save one.

The File menu has New workflow, Open .pdflow, Save and Save As, and Edit as JSON for when you would rather work on the file itself. It also lists ready-made workflows to start from, among them Weekly Social Posts, Appointment Reminders, Morning Brief, Review Responder and Competitor Teardown.

Workflow properties

Name and description

What the workflow is for.

Provider / model

The model every prompt and agent step uses unless the step picks its own.

Parameters

The inputs a run asks for, each with a type (text, number, yes/no, a list of choices, a date and more), a description, a default and whether it is required.

Guardrails

Checks on the run’s input and its output that deny, scrub, reroute or warn. The first one that matches wins.

Add node

Every workflow begins at Start and finishes at End. Add node puts any of these between them.

Prompd

A prompt, sent to the model.

Agent

A prompt that can use tools over several turns, with its own guardrails on the way in and out.

Input

Pauses the run and asks you for a value.

Guardrail

Has the model judge the output so far against a policy, and stops the run if it fails.

Mapper

Reshapes data with a template. No model call.

Branch

Takes the True or False path depending on an expression; the paths rejoin afterwards.

Loop

Repeats its steps for each item, a set number of times, or while a condition holds, up to a maximum (100 unless you change it), and collects each pass’s output.

Parallel

Runs branches side by side.

API

Calls a web address with the method, headers and body you give it.

Debug

Shows what passes through it and changes nothing.

Parallel branches

A Parallel node decides when it is done and what it hands on.

Wait for

Wait for all, First to finish, First success, or Quorum (continue once that many branches succeed).

Combine results

By branch name, as a list, or just the first result.

Max at once

How many branches run at the same time, 4 by default. A branch can also be given a timeout.

Expressions

Branch conditions are evaluated by a safe expression evaluator, never run as code. An expression that cannot be evaluated fails the step rather than being treated as true or false.

Run a workflow

1. Choose Run. If the workflow has parameters, fill in Run inputs first.
2. Watch each step in the Run pane as it finishes. Stop ends the run.
3. Read the Final output, or the reason the run failed, along with the tokens used, the estimated cost and the total time.

Inspect a step

While the run is on screen, each step can show its Prompt (the prompts exactly as sent), its Context (the values it was given) and its Settings (the model it used). Raw shows the run’s events instead of steps.

Inspecting is for the run you just watched. A run replayed later from Activity does not carry the per-step detail.

Past runs

Every run is recorded and appears under Activity, Runs, where you can replay it. See Activity: runs and sessions.

Read-only unless you say otherwise

A prompt or agent step can read your workspace and its own memory and nothing else. To give it more, attach Tools to the step and tick what it may use: Web search, HTTP request, Memory & key-value, Workspace files (read-only), Workspace files (write), Compile prompd, Registry search, Shell command, and the MCP tools on a connected sidecar.

What still asks

A tool you tick is one you have allowed, so it runs without asking. Shell command is the exception: it asks on every call, and it needs a connected sidecar started with --allow-shell.

Every tool call is still checked against the standard policy and recorded, and a run you start yourself honours the tools you have set to always allow. See Approvals, policy and Home.