Workflows

Workflows let you orchestrate multiple agents into pipelines — chaining them together so the output of one becomes the input to the next. This is how you build complex, multi-step processes: a document arrives, one agent extracts data, another validates it against policy, and a third writes the result to your system of record.

Each step in a workflow runs in its own agent context with its own competencies, memory, and PII boundary. This isolation is deliberate — it prevents data leakage between processing stages and lets you assign different trust levels to different steps.

What you'll find here

GuideDescription
Creating workflowsDefine workflows in JSON: agent references, variable passing, step configuration
Step modesControl execution flow: Sequential, FanOut, Collect, Conditional, Loop
Error handlingConfigure failure behavior: Fail, Skip, Retry — plus timeout management

A simple workflow

Each agent in the pipeline has its own competencies and PII boundary. Variables flow between steps explicitly — nothing leaks implicitly.

Common tasks

Step modes at a glance

ModeBehavior
SequentialSteps run one after another (default)
FanOutParallel execution across items
CollectGather FanOut results into a single output
ConditionalBranch based on a variable or previous output
LoopRepeat a step until a condition is met
  • Agents — each workflow step runs inside an agent
  • Competencies — agents in a workflow are shaped by their competencies
  • Data integration — workflows often read from and write to external systems