Skip to content
Talk to our solutions team

Workflows Operations

Workflow definitions are loaded per tenant. GET /workflow/config returns what is currently loaded — worth checking after a deploy, since a definition that failed to load produces no error at start time, only a failure when something tries to use it.

Instances execute on a pool of connected agents. Steps carry error handling and retry policy, so a transient failure in one step does not discard the instance’s accumulated state — the step retries, and the instance continues.

That is the important difference from a job: a failed step in a workflow does not mean starting over.

An instance is stuck. Usually it is waiting for a signal that nobody sent. Check /admin/instances for wait state and age before assuming a defect.

A step keeps retrying. A flaky integration inside an otherwise healthy process. The instance is fine; the dependency is not.

Queue depth grows. Not enough agents for the instance start rate.

A definition change lands mid-flight. Running instances continue on the definition they started with. Plan version transitions rather than assuming a deploy migrates in-flight work.

UseWhen
JobsFire-and-finish background work, on a schedule or trigger
WorkflowsBusiness processes that suspend, wait for humans, and resume
Data PipesMoving and transforming data in bulk
AI FlowTask DAGs including model calls; harnesses for agentic loops
SignalWhy it matters
Instances in wait state, by ageApprovals nobody is actioning
Queue depthNot enough agents for the instance rate
Cancelled vs completed ratioProcesses users abandon partway
Step retry rateA flaky dependency inside a healthy process
Definitions loadedA failed load is silent until something needs it

Wait-state age is the signal with business meaning rather than technical meaning. A rising median means your approvers are the bottleneck, which is a staffing answer, not an engineering one.