AI Workflows
A workflow is drawn once — agent steps, quality loops, human gates — then any card can follow it. This guide covers the editor, the step types, and how to read runs.
Open AI WorkflowsWhat is a workflow?
A workflow is a reusable delivery circuit: the sequence of steps a card travels between “started” and “shipped”. Each step is held by an agent, a rule, or a human. Where an automation reacts to a single event, a workflow orchestrates a whole journey — with branches, loops, and approvals. Draw it once, then run it on any card, in one project or across the workspace.
Drawing the circuit
The editor is a canvas: drop steps, wire their outputs, and the circuit reads left to right. Green paths are the nominal journey; red paths are rejections and returns. Each step is configured in place — which agent, which instructions, which exit condition. “Test” plays the circuit dry before you activate it, and every workflow keeps its run history.
Step types
Circuits are composed from a small set of bricks — the same ones you see in the run view:
Start / End
The circuit’s endpoints: the card comes in, the card ships out.
Agent step
An agent (Atlas, Forge…) executes the step’s mission with its instructions.
Review & score
A reviewer agent grades the work — for example a score out of 20.
Condition
A branch on a run variable: score, severity, card type.
Loop
Send work back to an earlier step, with a cap (e.g. 3 fixes max).
Human gate
The run stops and waits for an explicit approval from a person.
Notification
Ping a channel or a person — delivery, alert, review request.
Board action
Move the card, change a status, create a follow-up card.
Quality loop & human gate
The most useful pattern combines three bricks: one agent builds, a reviewer agent scores, and a condition sends the work back for fixes while the score stays under your threshold — with a capped number of rounds so it can never loop forever. Once the bar is met, a human gate guards the exit: nothing ships without your sign-off.
That is the “Iterative quality control” circuit the home demo plays: quality improves inside the loop, and the final call stays human.
Interactive demo
The circuit below plays right here in the page — steps lighting up, scores, a fix loop, an approval. No data is sent anywhere.
For advanced workflows: pipelines
When a mission needs several steps - produce, review, fix, request approval, notify - a pipeline can orchestrate the sequence. Powerful, but not required to launch your first AI mission.
Pipeline
Checkout release pipeline
Ship checkout without escalating the whole routine to a human.
Execution log
Run started
State
Step 1 / 7
Route
Production
Approval
None
What the pipeline just did
Ship checkout without escalating the whole routine to a human.
Checkout release pipeline
Produce, review, correct if needed, request a human green light, then notify automatically.
Pipeline
Ship checkout without escalating the whole routine to a human.
Pipeline editor
The pipeline picks the next step on its own.
01
Produce
Product AI builds the first version.
02
Review
Review AI scores the output and surfaces risks.
03
Score >= 8?
The pipeline decides whether the work can pass.
04
Payment green light
Only one decision reaches the product lead.
05
Loop back
The pipeline launches another pass if the score is too low.
06
Correct
Fix AI applies changes and sends the work back to review.
07
Notify
Slack and release docs update without manual work.
Following runs
Every launch creates a run: the run view shows the circuit with the current step, the “Live” journal of events, and the counters (active, finished, blocked). A waiting human gate surfaces as an approval to give — from the run view or from the card itself. History stays attached to the workflow, so you can compare runs and see where things snag.