AI automations your agents set up

AI automations (formerly Crewdle Forge) is a tool your agents use, not a builder you open. When a job needs the same AI steps to run the same way every time, like reading an uploaded document and returning structured fields, classifying incoming requests, or answering from your own documents inside an app, an agent designs that pipeline as a workflow, tests it, and deploys it for you.

Agent or workflow?

A hired agent decides for itself: it watches, judges and acts. A workflow does not improvise: it runs a fixed sequence of steps each time it is called, which makes it predictable, cheap to run at volume, and easy to audit. Use a workflow when an app, a website form or another system needs one reliable AI step. Agents reach for this tool when they spot that kind of job, and you can ask for one directly.

What a workflow is made of

A workflow is a pipeline of blocks. Each block does one thing: it receives the input, reasons with a model and instructions, looks things up in a data collection (your documents, for grounded answers), calls a connector, transforms data, or returns a result. A block can force a structured output when your app needs a predictable shape. The agent picks the blocks, wires them together and chooses the models.

Give an agent the tool

How to give an agent the AI automations tool

1. Your org admin keeps the tool (under Crewdle tools) turned on in Admin → Tools, optionally limited to some roles.
2. Open the agent, then its Tools panel, and click Connect on AI automations. Developers on a website or app project usually get it with the project. Tool grants are set at hire time, so re-grant the tool if an older agent does not see it.
3. Describe the job: what goes in, what should come out, and where it is used. For example: "Our quote form should read the uploaded floor plan and return the room count and square footage as JSON."

How the agent ships it

1. Design. The agent proposes the steps and the models, and builds the workflow.
2. Test. It runs the workflow on sample inputs and reads the run traces to check each step.
3. Deploy. It asks before publishing: a Publish an AI automation card appears in the conversation, and nothing goes live until you approve. Workflows are versioned, so a later change can be rolled back.
4. Connect. Your website or app calls the deployed workflow. When a Crewdle agent is building that app, it wires the call in for you.

Here, the Developer built a workflow that sorts the notes visitors leave on a waitlist form and tested it on four samples, including a trick note that tries to give it new instructions:

Credentials the workflow needs come from your team's secrets vault: the agent binds a stored key to the workflow without ever reading it. See Security and privacy.

Private mode for sensitive data

For confidential documents, a workflow can run in private mode. A private run leaves no trace in the run logs, and on OpenAI models the provider is asked not to store the request. Zero data retention requires every model in the workflow to be an OpenAI model, so tell the agent up front that the workflow must run private. Debug with private mode off first, since private runs cannot be inspected afterwards.

Find and manage your workflows

Workflows the agent built appear under Deliverables in its Files panel, with a menu to download the workflow definition or delete it. To change one, ask the agent: it updates the workflow, re-tests it and deploys a new version.

What it costs

Building and testing a workflow is billed as the agent's task. Each later run of the deployed workflow is billed in credits per run, from the model usage of its steps, on your one Crewdle bill. The Usage dashboard in Admin shows workflow spend. See Credits and billing.