Processes and execution

A process is a nested workflow that coordinates people, AI, scripts, integrations, and data. Steps run sequentially along the selected path.

Execution lifecycle

1. A user, an AI assistant, or another process starts the process.
2. Required start variables are validated against their JSON Schemas.
3. Steps run in order. Branches can select a different path.
4. A task can pause execution until its completion requirements are met.
5. A Return step ends execution and exposes selected output variables.
6. Process variables are deleted after execution ends.

Step types

Every step name must contain 1–32 characters.

Script

A Script step can read or write selected variables, access tables, and execute selected processes. Its program must finish; scripts cannot be long-running services.

Agent

An Agent step can read or write selected variables and call selected processes. Optional terminal commands run in the selected sandbox; the AI agent itself does not run inside it.

The agent’s final message is logged. It does not update a variable unless the agent explicitly writes that variable.

Example Agent prompt:

Read the instructions from $request. Use the selected sandbox terminal to
complete the request, then save the result to $result.

The Agent step must grant access to $request, $result, and the sandbox terminal required by the prompt.

Task

A Task assigns the same action to users resolved from a user expression.

• 1 user wins: execution continues after the first submission.
• All users are needed: execution continues after every assigned user submits.
• An optional deadline can end the waiting period.

Task output variables must be arrays. Each submission contributes one item to every output variable:

await ailaflow.submitForm({
  answer: ['Approved']
});

Two submissions produce an array such as ['Approved', 'Rejected']. Matching indexes across all task output variables belong to the same submission.

Task metadata can identify who submitted each item and when:

{
  status: 'completed',
  items: [{ time: 1770000000000, userName: '@robert' }]
}

Notification

A Notification resolves a user expression, stores one notification per matched user, forwards it to an active default chat when available, and continues immediately.

Branch

A Branch reads a string variable and executes the branch whose name exactly matches its value. Define a branch for every expected value.

Return

A Return step stops execution immediately, including from inside a branch. Selected variables become the process result.

Variables and schemas

Every variable has a JSON Schema. Define complete object properties and array items; avoid an unspecified schema such as { "type": "object" }.

Start variables are required process inputs. A caller must provide all required values with matching types.

Use variables for execution state:

• Form input
• Intermediate calculations
• Task submissions
• Branch selectors
• Process results

Persistent data

Process variables disappear when execution ends. Use tables for data that must remain available.

Tables use string primary keys and store JSON-compatible values. Every row also contains _id and an automatically managed _updatedAt timestamp.