No-Code Workflows (DAG Automation)

Overview

Workflows let you chain steps into an automated pipeline — a DAG (directed acyclic graph) of nodes — without writing code. A workflow can generate content, call AI capabilities, hit external APIs, branch, loop, wait, get approvals, and publish results. Each step runs with its own timeout, budget, and error policy so a single flaky step can't run away or hang the whole run.

Node kinds

Workflows support these step types:

1. Capability — call an AI capability (chat, image, TTS, translate, etc.)
2. Tool — invoke a registered tool/function
3. HttpRequest — call an external HTTP endpoint
4. Branch — route based on a condition
5. Loop — iterate over a collection
6. FanOut / FanIn — parallelize and rejoin
7. Transform — set/reshape variables
8. ArtifactSave — persist a transient asset to your CDN
9. Publish — push to a channel (webhook, markdown artifact, etc.)
10. SubWorkflow — call another workflow as a step
11. HumanApproval — pause for an approval
12. Wait — delay
13. Webhook — POST to an external hook

Per-step controls

Each node carries:

• Timeout — how long the step may run before it's cut off.
• Error policy — what happens on failure: fail the run, retry, or continue/ignore. For example, set a notification step to ignore so an outage there doesn't fail the whole run, and set a fetch step to retry for transient upstream blips.

Variables, secrets, and artifacts

• Variables thread data between steps. Steps read prior outputs with {{vars.name}} interpolation.
• Secrets hold credentials a workflow needs (API keys, webhook secrets) without hardcoding them into the spec.
• Artifacts are outputs you persist — ArtifactSave re-hosts transient provider URLs to permanent Bunny CDN URLs so links don't expire.

Cost

There is no separate workflow execution fee. A run is billed as the sum of every step's BillingRecord — exactly the same per-call billing as running each step by hand. Steps that make no AI call (HTTP, transform, wait, webhook) cost nothing. Sub-workflow costs roll up into the parent run's total.

Running and monitoring

Kick off a run, then watch it on the workflow runs page. Each run records its steps, their outcomes, and total cost. Failed steps surface their error per the step's error policy.

Tips

• Give external/notification steps a forgiving error policy (retry or continue) and core steps a strict one.
• Use ArtifactSave on any generated media you want to keep — provider URLs expire.
• Extract reusable subgraphs into a SubWorkflow and call it from multiple parents.
• Test a flow with cheap echo/transform steps before wiring in costly AI capability steps.