Custom workflow agents
BetaDefine your own multi-step agent — AI calls, data queries, HTTP calls and notifications, run in sequence.
Plan availability. Building your own agent is gated on a feature key, and this is not switched on in any of our four published plans — Starter, Pro, Business and Enterprise all have it off — so a subscription will not unlock it. It has to be enabled for your organisation by arrangement. Ask us before you plan around it.
- ✓Executable step types — AI call, database query, API call, condition, wait, email, in-app notification, outbound webhook
- ✓Six trigger types definable; manual and email are the two that dispatch
- ✓Test run against sample input, rolled back so nothing persists
- ✓Numbered versions on every save
What it is.
Build an agent from steps that execute in order: an AI call, a database query, an outbound API call, a wait, a condition that stops the run when it does not hold, and delivery steps for email, in-app notification or a signed webhook. Save it and the definition is versioned; test it against sample input and the run is rolled back so nothing is written.
Two limits worth stating rather than discovering. Steps run in sequence — a condition that fails stops the run, it does not branch to an alternative path. And two step types you can save, loop and human approval, are not executed by the engine: they will validate when you save and fail when they run.
Three steps.
End to end.
Manual run or an email inbox address. Cron, webhook, contract-upload and compliance-alert triggers can be saved but are not dispatched.
AI call, database query, API call, condition, wait, then email, in-app or webhook out. Data passes from step to step.
A test run executes the real engine against your sample input and rolls the transaction back. Saving creates a new numbered version.
What you get.
- ✓Executable step types — AI call, database query, API call, condition, wait, email, in-app notification, outbound webhook
- ✓Six trigger types definable; manual and email are the two that dispatch
- ✓Test run against sample input, rolled back so nothing persists
- ✓Numbered versions on every save
- ✓Per-run record with each step's result and any error
- ✓Publish an agent to the agent marketplace, or install one another organisation on this product published
Quick answers.
No. Steps run in order; a condition that does not hold ends the run rather than routing elsewhere. Success and failure targets can be declared on a step and the engine does not act on them.
Versions are kept and numbered, but there is no endpoint to activate an earlier one. Rollback means editing forward.
Both have config schemas and neither is implemented in the engine — a run reaching one fails. Use Deal Desk for a real approval gate on a contract.