AI workflows
A webhook goes in. Finished work comes out.
Design the process on a canvas, trigger it from anything that can make an HTTP request, and let agents do the work. Pause for a human decision when it matters.
Nodes
Thirteen building blocks
A workflow is nodes and the connections between them. Drag them onto the canvas, wire them up, and the shape of the process is the diagram rather than a description of one. Connect one node to several and those branches run in parallel.
Entry point
Invoke an agent
Call a model directly
Call a tool directly
Apply a skill
Create or update a task
Branch on a test
Bring branches back
Wait a while
Ask a human
Block until done
Run a batch job
Post to Slack
Triggers
Start it from anywhere
Give a workflow a webhook trigger and anything that can make an HTTP request can start it — a GitHub event, a Slack command, a form submission, a device, another workflow. The request body becomes the workflow’s variables.
POST or GET. Returns an execution ID immediately and runs in the background.
Cron expressions for nightly sweeps and weekly audits.
Run it yourself from the dashboard while you are building it.
Approvals
Stop and ask a person
Full automation is the wrong answer for anything consequential. An input node pauses the workflow and waits for a human — a free-text answer, a choice from a list, a number, a yes or no, or a file.
The execution holds its state while it waits. When the answer arrives the workflow resumes from that node with the input available to everything downstream. Set a timeout if waiting forever is not acceptable.
This is how you put an agent in front of a decision that needs sign-off without giving it the authority to make that decision.
Orchestrations
Run agents over a backlog
A workflow handles one job. An orchestration handles thousands. Give it the work — a fixed list, a task query, an API endpoint, or a prompt that generates the items — then set batch size, concurrency, and a stop condition.
Stop when everything is done, after N iterations, when a model judges it complete, or when you say so
- Pause, resume, and cancel while it runs
- Read the batch history afterwards
Durable — it survives restarts rather than starting over
Versioned by default
Each time you publish a workflow, it creates an immutable revision. You can list them and restore one that worked.
Workflow definitions are JSON through the API, so they can live in version control alongside the rest of your infrastructure if that is how your team works.
Agents that coordinate
An overseer agent can create a task list, assign tasks to other agents, watch their progress, and decide what happens next — reassigning work that stalls, adding tasks it did not anticipate.
Tasks persist across conversations, so the state of the work outlives any single interaction.
What people automate
The pattern is a trigger, some agent work, and a place for the result to land.
Pull request review
A GitHub webhook starts the workflow, a review agent reads the diff, and the result lands as a review comment.
- Webhook triggered
- Runs per PR
- Comments in place
Scheduled audits
A cron trigger runs a sweep overnight and opens an issue with what it found. Same check, every week, no one has to remember.
- Cron triggered
- Consistent checks
- Issue per run
Bulk content work
An orchestration runs an agent across a backlog of documents with controlled concurrency, and you watch the progress.
- Batch processing
- Controlled concurrency
- Pause and resume
Automate the work around the work
Build a workflow on the canvas and trigger it from whatever you already use.