Skip to main content
Workflows let you create AI agents that handle complex, multi-step work automatically. Define what you want done, add the tools and context the AI needs, and let it run—on a schedule or whenever you trigger it.

​
What can workflows do?

A workflow is basically an instruction manual for an AI coworker. You describe the goal, list the steps, and the AI figures out how to execute each one. Some examples:
  • Weekly reports: Pull data from connected apps, summarize it, post to a project channel
  • Content pipelines: Research a topic, draft an article, create social posts
  • Customer follow-ups: Check for new support tickets, draft responses, flag urgent ones
  • Code reviews: Scan recent commits, note potential issues, create tasks for fixes
The AI can use any tools you’ve connected—GitHub, Gmail, Slack, Shopify, and dozens more.

​
Creating a workflow

1

Create a workflow page

In any project’s Resources section, click + New and select Workflow.
2

Write your goal

Describe what you want the workflow to accomplish. Be specific—the clearer you are, the better the results.
3

Add steps

List the steps as a numbered list. Each step should be a single action the AI can complete.
4

Connect integrations

If your workflow needs external tools (GitHub, email, etc.), add them from the configuration panel.
5

Add context

Include any pages, files, or background info the AI should reference while working.

​
Anatomy of a workflow

A workflow page has three main parts:

​
Configuration block

At the top of the page, you’ll see settings for:
SettingWhat it controls
AI modelWhich AI model runs the workflow (affects speed and quality)
ScheduleWhen to run automatically (manual, hourly, daily, weekly, monthly)
IntegrationsWhich connected apps the workflow can use

​
Goal section

A plain-language description of what the workflow should accomplish. Write this like you’re explaining the task to a new team member. Good goal:
Check our GitHub repository for any open PRs older than 3 days. For each one, post a reminder in the #dev channel asking the author to address it or close it.
Vague goal:
Help with GitHub PRs.

​
Steps section

A numbered list of actions. Each step should be:
  • Specific: One clear action
  • Achievable: Something the AI can actually do with available tools
  • Sequential: Order matters—later steps can depend on earlier ones
1. Fetch all open pull requests from the main repository
2. Filter to PRs that have been open more than 3 days
3. For each PR, look up the author's Slack handle
4. Post a message in #dev tagging the author with a link to their PR
5. Create a summary of all PRs you messaged about

​
Running workflows

​
Manual runs

Click the Run button in the workflow header to start immediately. You’ll see real-time progress as the AI works through each step.

​
Scheduled runs

Set a schedule in the configuration block:
ScheduleWhen it runs
HourlyEvery hour on the hour
DailyOnce per day (you pick the time)
WeeklyOnce per week (you pick the day and time)
MonthlyOnce per month (you pick the date and time)
Scheduled workflows run automatically. You’ll see results in the workflow’s run history.

​
Watching a workflow run

When a workflow executes, you can watch it in real time:
  • Step timeline: See which step is currently running
  • Tool calls: Watch the AI interact with connected apps
  • Progress indicators: Track completion percentage
  • Token usage: Monitor how much AI capacity the run uses
The AI shows its work—you can see exactly what it’s doing at each step.

​
Review gates

Sometimes you don’t want the AI to barrel through every step without checking in. Add review gates to pause execution and wait for your approval.

​
How review gates work

  1. The workflow runs until it hits a review gate
  2. Execution pauses and you get a notification
  3. You review what the AI has done so far
  4. Choose to approve (continue), reject (stop), or regenerate (try again with feedback)

​
When to use review gates

  • Before sending external communications
  • Before making changes that are hard to undo
  • When the AI needs to make judgment calls
  • For high-stakes steps where you want human oversight
Review gates are optional. Simple, low-risk workflows often don’t need them.

​
Workflow results

After a workflow completes, you get:
  • Output summary: What the workflow accomplished
  • Tool call log: Every action taken, with parameters and results
  • Resources created: Any tasks, pages, or other items the workflow made
  • Run duration: How long execution took
  • Token usage: AI capacity consumed
All runs are saved in the workflow’s history, so you can review past executions anytime.

​
Integrations in workflows

Workflows become much more powerful when connected to external tools. When you add an integration, the AI can:
IntegrationWhat workflows can do
GitHubRead/create issues, PRs, comments, check repo status
GmailSend emails, search inbox, draft replies
SlackPost messages, read channels, send DMs
LinearCreate/update issues, manage projects
ShopifyCheck orders, update products, manage inventory

Set up integrations

Connect apps to enable them in workflows.

​
What workflows create

Workflows can generate resources as they run:
  • Tasks: Create task records for follow-up work
  • Pages: Generate documents with findings or reports
  • Messages: Post updates to project channels
  • Calendar events: Schedule meetings or reminders
These show up in your workspace like any other resource, tagged with which workflow created them.

​
Troubleshooting workflows

Check the error message in the run log. Common causes:
  • Missing integration permissions
  • API rate limits on connected services
  • Step instructions that are unclear or impossible
Try rephrasing the step or checking your integration connections.
The AI interprets your instructions, so vague goals lead to unpredictable results. Review your goal and steps:
  • Is each step specific enough?
  • Did you provide enough context?
  • Are there edge cases you didn’t account for?
Complex workflows with many steps take time. To speed things up:
  • Combine steps that can be done together
  • Use a faster AI model for simple tasks
  • Remove unnecessary steps
Check that:
  • The schedule is set correctly in configuration
  • The workflow isn’t paused
  • There are no integration connection issues

​
Tips for better workflows

Build a basic version first and make sure it works. Add complexity gradually—it’s easier to debug a simple workflow than a complex one.
“Check GitHub for issues” is vague. “Fetch open issues labeled ‘bug’ from the acme/api repository” is specific.
Run the workflow manually a few times before setting up automated scheduling. Make sure it behaves the way you expect.
Any step that sends external communications or makes permanent changes should probably have a review gate.
The more relevant context you provide (pages, examples, background info), the better the AI can execute your instructions.

​
Next steps

AI capabilities

See everything AI coworkers can do.

Integrations

Connect external apps for workflows.