Appearance
Routine workflow map
Use this page to choose the smallest routine construct that fits the business process. Then give the outcome to Claude or Codex through the Agents MCP; the coding agent can read the complete YAML contract, inspect the account, validate the draft, and run a bounded preview. You do not need to assemble YAML by hand.
This is the implementation map for the routine-first operating model: the routine owns repeatable orchestration, while agents, skills, connectors, knowledge, and models provide capabilities to its steps.
Start with the outcome
| The process needs to… | Use | Read next |
|---|---|---|
| Run one or several steps in order | Agent steps or sequence | First routine · Sequence |
| Run independent work at the same time | parallel | Sequence and parallel |
| Repeat until a bounded quality condition is met | loop | Loop |
| Choose a branch from a test | condition | Condition |
| Let a model select the best route | router | Router |
| Apply the same workflow to a list discovered at run time | foreach | Process a list |
| Stop cleanly when there is no useful work | early_exit | Early exits |
| Run on a clock | schedules | Scheduling |
| Run when an external event arrives | Trigger | Triggering |
| Balance quality and cost step by step | Per-step model | Model selection |
Large collections
Use foreach when the server can return records in bounded pages and the routine should perform the same work for each item. The source step discovers a page or list; the item workflow handles individual records; a final step aggregates ordered results and partial failures. Keep pagination and the total-record loop explicit in the routine or skill script instead of asking one server call to export the entire dataset.
The authored limits are deliberately finite:
| Limit | Maximum |
|---|---|
Resolved foreach items | 10,000 |
foreach item concurrency | 100 (default 10) |
| Combined item-worker and nested parallel width | 100 Agent runs |
| Loop iterations | 100 |
| Total authored workflow nodes | 64 |
| Nesting depth | 8 |
For a worked pattern with stable identifiers, bounded concurrency, ordered results, and partial-failure handling, follow Process hundreds of records.
Authoring path
- Open an Agent's Tasks tab and choose Connect Claude or Codex.
- Ask for a routine in terms of the business outcome, record source, effect boundary, completion criteria, and desired schedule or trigger.
- Let the client call the authoring-context and inventory tools before drafting.
- Review the canonical YAML and validation result.
- Run a small, tool-free preview; test real connector reads separately when required.
- Keep schedules and triggers disabled until the routine and any external effects have been reviewed and tested.
Preview is not a side-effect sandbox
A routine preview proves orchestration and model behavior without exposing configured tools. It does not prove that an external read or write will work. Test connector access separately, and follow the current approval-state warning in the YAML reference before enabling any message, call, publication, deletion, or live record update.
Keep the complete contract available
The Routine YAML reference is the authoritative field, validation, cap, and error reference. Its raw Markdown version is available at https://docs.mychatbot.app/agents/routine-yaml-reference.md for agent clients.
See also
- Agents Platform operating model — why the routine is the central production entity
- Build routines with Claude or Codex — complete authoring and staged-activation flow
- Custom routines — lifecycle in the app
- MCP surface map — choose and connect the Agents MCP
- Routine YAML reference — complete canonical schema