Products and gates

A worker's completion report is a report, not a verdict. Products describe promised graph state; gates decide whether finished work can pass. All declared products must hold and all gates must pass.

Products

produces names state the work promises to create, such as a resource in a particular state. Fractal checks the promise against the graph. Until it holds, the step remains in verification with the unmet conditions visible.

Gates

• Graph predicates check recorded facts.
• Mechanical gates run a command with a time limit.
• LLM gates ask a model to judge with explicit limits and tools.
• Human gates ask a named person to decide.

A graph or human gate can remain pending until there is enough evidence. A failed gate is not success just because the worker reported completion.

Human review

Inside a mission, place this after a worker's prepare step:

step "approve" {
  agentless
  depends-on { step "prepare" completed }
  gate "The operator approves" type="human" {
    reviewer "person/ada"
    question "Is the result ready to use?"
  }
}

The request appears in the reviewer's Now view with its available actions. The agent remains free for other ready work while the review waits.

Declare checks before execution so the acceptance standard is visible to everyone. Read the mission language reference for exact gate syntax and Missions in practice for a complete review flow.