# Orchestrate (PoC) — Claude Code

Botize method — `claude_code.orchestrate` (action).

Iterative orchestration step. Calls Claude Code with the orchestrator schema (done/message/next_agent/reason) and pauses until the response (or a manual stop) arrives. Multi-agent delegation is not implemented yet: when 'done' is false the task ends with an informational text.

This is an action: it runs as a step of a task. Several can be chained.

## Fields you set

The field's internal name, its type and the values it accepts. These fields accept `{{...}}` tags from earlier steps.

- `mb_switch` *(boolean)* — Rules
- `session` *(string)* — Session name
- `message` *(string)* — Initial message
- `task_ids` *(string)* — Available agent task IDs
- `max_iterations` *(number)* — Maximum iterations
- `mb_outputvars_switch` *(boolean)* — Customize tags

It also accepts repeatable field groups, with these prefixes: `mb_filters_*`. These are composite components (buttons, filters, keyboards); the trailing number is set by the user as they are added.

## Output variables

Data this step makes available. Reference them in later steps with the double-brace tag.

- `{{text}}` *(text)* — Final message. Final message produced by the orchestrator. Present when done=true or stopped=true.
- `{{cost_usd}}` *(text)* — Cost (USD). API cost in USD for the last call. e.g. 0.0032
- `{{duration_ms}}` *(int)* — Duration (ms). Total response time of the last call in milliseconds
- `{{session_id}}` *(text)* — Session ID. Internal Claude session UUID (for resuming)
- `{{job_id}}` *(text)* — Job ID. Worker job identifier of the last call
- `{{ok}}` *(boolean)* — Success. true if the call succeeded, false otherwise
- `{{done}}` *(boolean)* — Done. 'true' when the orchestrator finished. Always 'true' at this output (delegations re-pause the task instead of emitting outputs).
- `{{next_agent}}` *(text)* — Next agent (last delegation). Empty at this output. Surfaced only via Pushover before re-pausing.
- `{{payload_json}}` *(text)* — Payload JSON (last delegation). Empty at this output. Surfaced only via Pushover before re-pausing.
- `{{reason}}` *(text)* — Reason. Brief reasoning behind the orchestrator's last decision.
- `{{permission_denials}}` *(text)* — Permission denials. JSON array of tools Claude tried to use but were blocked. Empty when there were none.
- `{{stopped}}` *(boolean)* — Stopped manually. 'true' when the orchestration ended because the user clicked a stop link.
- `{{iterations_done}}` *(int)* — Iterations done. Number of callback iterations completed before finishing.

## Account

Needs a connected Claude Code account in Botize: https://botize.com/en/app/claude_code/authenticate

## This app

Every trigger and action for Claude Code: https://botize.com/en/app/claude_code.md

## How to build this

An AI assistant (Claude, ChatGPT, Cursor…) can create and edit these tasks in the user's account through the Botize MCP connector, included in every plan:

    https://webhook.botize.es/mcp

- Connection guide: https://botize.com/en/mcp
- Agent map: https://botize.com/llms.txt
- New account: https://botize.com/en/join (500 free operations, no card)

---
Machine-readable summary for AI assistants. Human version: https://botize.com/en/method/claude_code/orchestrate