Workflows
Build repeatable processes on a visual canvas, with AI steps, people, logic and tools. Run them by hand, on a schedule, or when another system calls them, and watch every step.
A workflow is a process you want done the same way every time. Each one has a goal, a way to tell whether it succeeded, inputs and outputs, a set of connected steps, and optional triggers that start it. Steps can be AI turns, people, logic or tools.
Workflows live in the Agency window on the Workflows tab. Open it from Agency in the sidebar.
The workflow library
Workflows sit under category headings. Uncategorized is the default and cannot be deleted. Rename a category with the pencil on its heading or a double-click, delete one with the bin (its workflows move to Uncategorized), and drag a card to move it. The toolbar has search, a category filter, a cards or table view, Import Workflow JSON, and a button for a new category.
Each card shows the name, Goal, Success, Status and Next run. Badges add the category, whether it is a draft or published, the number of steps, and how much of its success test is covered. Live badges show running, needs you and waiting counts.
| Status | Meaning | Next run shows |
|---|---|---|
| Draft | Never published. | Publish it to run |
| Manual | Published, starts only when you run it. | When you start it |
| Active | Published with a schedule or an event trigger. | A time and countdown, Running now, or When its trigger fires |
| Paused | Automatic starts are on hold. | None while paused |
Card buttons:
- Open, the first button, in the workflow's own colours, opens the editor.
- Pause and Resume appear on Active and Paused cards. They hold or restart only the automatic starts; you can still run the workflow by hand.
- Info shows everything about it: steps, criteria, triggers, inputs, outputs, policies and platforms.
- Edit changes the name, description, category, memory, color, cover and icon without touching the canvas.
- Delete asks first.
Create a workflow
- Press New Workflow, give it a name, and press Create Workflow. The canvas opens with one Start step.
- Press + (Add a step) and choose Agent Task. Fill in Expected result, Action and the Objective.
- Drag from Start's exit dot on the right to the new step's dot on the left.
- Add an End Branch step and connect it.
- Open ⋯ → Goal and success and describe the goal and how success is measured.
- Press Save, then Run, and read the result in the Output tab.
The canvas
The header holds the back arrow, the workflow's name, Run, Publish, Save, ⋯ and close. ⋯ opens Goal and success, Policies, Revisions, Runs and Download. A key button also appears when the workflow can be started by an outside system.
- Move around: drag empty space to pan; the mouse wheel zooms. The controls in the bottom right are Zoom out, Readable zoom, Zoom in, AI chat, Full overview and a switch between compact and detailed cards. The minimap shows where you are.
- Steps: each card shows an icon, a name and a one-line summary. Show settings (or a double-click on its header) opens its fields. Hover a card for step details, rename and remove.
- Connections: drag from a dot on the right of one step to a dot on the left of another. Click a connection to delete it, and press Delete to remove the selected step.
- Exits: a step can have several named exits, such as success and error, approved and rejected, pass and fail. For Branch, Router and Do at once, you name the exits yourself.
- Missing fields: steps with missing required fields are flagged, and required fields carry a star.
Canvas edits have no undo. Press Save often, and use Revisions to go back to a published version.
Steps you can add
Press + for the Add a step panel. Search every step... finds any step. Quick steps are Start, Agent Task, Branch, End Branch, Human Review, Loop, Notify and Region. The rest sit in folders:
| Folder | Steps and what they do |
|---|---|
| Core | Agent Task runs one full AI turn in its own session. Think, Understand and Execution Plan are focused reasoning passes. Team runs one of your teams. End Branch ends a path. |
| Flow Control | Branch and Loop. Do at once runs branches in parallel and Wait for branches joins them (all, any or first). Wait pauses for a time, until a time, or for an event. Run a workflow calls another workflow. Region groups steps as one unit that fails and cleans up together. Decision takes the first matching rule. Transform. Human Review waits for approve or reject. Judge gives an AI pass or fail against a rubric. Router lets the AI pick one of your named branches. |
| AI & People | Ask pauses and collects answers from a person in plain language. |
| Data & Memory | Memory Query reads memory without an AI turn. Remember writes to memory. |
| Actions | Notify sends a fixed message. Deliver Result sends the final result through a channel. Run Code runs a shell command. HTTP Call calls an outside address. Document creates or transforms a document. |
| Tools | The product's own tools as direct steps, grouped by job, such as search and read, social research, communication and create. Expand a tool to pick an action. |
The i button on a step explains what it does, its settings, what it gives you, and when it counts as done.
Platforms
The Platforms button on the canvas lists steps for 105 outside platforms, from CRMs and spreadsheets to payments, databases, cloud services and social networks. Each platform step needs that platform's key on your instance. A platform without its key shows not connected: you can still place its steps, and they run once the key is added. Read-only research on Reddit, GitHub, Discord, Instagram, TikTok, Telegram, X and YouTube is also under Tools in the step panel. Integrations shows every platform with its logo.
Step settings and policies
Every step has an icon, an Expected result and an Action. AI-judged steps add Done when and Evidence. An Agent Task also picks the agent (Nova or a specialist), its modes, the objective, memory, tools and the shape of its output. When a field uses an earlier step's result, the canvas shows it in words, such as "the topic you provide".
Each step can carry its own Policies, for example who approves it, how many times it may retry, what happens on failure, its limits, and what it may touch. By default only high-risk steps ask a person to approve.
Goal, success and triggers
⋯ → Goal and success opens the Outcome panel:
- Goal, Achieved when, Measured by, Fallback and Evidence.
- Inputs the workflow takes (name, type, required, default) and Outputs it returns.
- Triggers: Manual (the Run button), Schedule, and Signed event (another system starts it by event name, such as
lead.created). Each trigger has its own on or off switch.
The schedule dialog offers Interval, Hourly, Daily, Weekly, Monthly, Fixed hours and Custom, with a live summary. Schedule changes take effect after you publish.
Workflow policies
⋯ → Policies sets rules for the whole workflow:
| Policy | Options |
|---|---|
| Memory | Its own workflow memory, the memory of the chat that started it, or the shared default memory |
| Source material | Inline text, or preserved sources |
| Run | The live draft, the published version, or a safe test |
| Publishing | Ask, When valid, or Nova builds |
| Outcome verification | Required (default), best effort, or off |
| If verification is unavailable | Pause (default), fail, or continue |
| Compensation | Manual, on failure, or none |
| Run history | How many days to keep runs (default 14) |
Publish and automate
Publish turns the draft into a version that schedules and events can start. A draft that is not complete cannot be published. After publishing, a scheduled workflow's card turns Active and shows its next run, and its schedule appears in the Scheduler.
To let another system start a workflow, add a Signed event trigger, publish, then use the key button to create its key. You can rotate or revoke the key there. A paused workflow refuses events.
Run and watch
Press Run to start the workflow now. If a required field is missing, the step is flagged and nothing starts. By default Run uses the draft that is open, including unsaved edits. A Safe test run holds back outside effects; a normal run does not.
The run panel shows the state in plain words, such as Working, Waiting for you, Waiting until the right time, Finished or Could not finish. It also shows how many steps are done, the elapsed time, and Stop. It has three tabs:
- Steps: a timeline of every step with its time, result and whether it met its success test.
- Output: the final result. It opens on its own when a run returns values.
- Activity: a plain-language log.
Click a step to see its result, attempts and setup; agent steps add Open step chat. When a step waits for a person, its row shows Approve and Reject, and the card shows a needs you badge. When a run waits for missing details, the Nova panel opens and asks for them, and the run continues once you answer. If something fails, you get one plain sentence explaining what happened.
Revisions, runs and backups
- Revisions lists published versions. Click one to restore it into the draft.
- Runs lists past runs, newest first, with their state, time and outcome. Click one to load it.
- Download saves the definition, the definition with revisions, or a full archive. Credentials and signing keys are never included.
- Import Workflow JSON in the library always creates a new private draft.
Build with Nova
The AI chat button on the canvas opens a Nova panel. Describe the outcome you want and Nova rewrites the canvas and saves the draft. From the main chat you can also ask Nova to list, build, change, test, publish, run, pause or delete workflows. Publishing, restoring, running, cancelling and resuming ask for your approval first.
Requirements
- AI steps (Agent Task, Judge, Router, Think, Understand, Execution Plan) need a working model.
- Platform steps need the platform's key.
- Delivery channels other than In the app must be set up in Settings; until then they show Setup required.