AI task

This is the step that runs an assisted task on Imaginne's engine. It can use tools and skills, and it is the only step whose result is not deterministic — which is why it is also the one that most needs to be declared well.

Four questions, in this order

The panel is organized around them:

1. What should the AI do? — the objective.
2. Which data does it use? — the input data.
3. May it look up information or use resources? — the tools.
4. What should the answer look like? — the expected output.

And it closes with the question that decides everything: how will this step be considered complete?

Objective and instructions

Field	Use
Objective	What you would ask a person to do. Required
Instructions	How to do it: tone, format, constraints. Optional

Insert data from earlier steps with the Data button — never by typing a technical path.

Instructions and data are different things

Instructions are what to do. Data is the content handed over to do it with.

• A value inserted into the middle of the objective becomes text inside the request.
• An item in the input data list is handed over as structure, with a name.

When the same value appears in both places, Studio warns you. It is not an error — but it has to be a decision, not an oversight.

Tools

Mode	What it means
None	Reasoning over the data handed over, and nothing else
Only the ones I choose	You select from the list of allowed resources
Any allowed in the Agent Space	The AI chooses among everything the space allows

You can mark a tool as required: without it, the step fails — and it fails before spending the model, when the unavailability is known in advance.

Looking up information is not switched on for you

Enabling search costs money and changes the result between runs. When Studio notices that the objective calls for a search or a system lookup, it suggests — the decision belongs to whoever is building.

Skills

The step can invoke the organization's skills. They must be allowed in the space and permitted by the organization's policy in Imaginne. Both conditions apply together: allowing something in the space does not bypass the policy.

Expected output

This is where the difference lies between a verifiable step and a step that merely "answered".

Format	When to use
Text	The answer is the text itself, and will be used as such
Fields	You declare the name, type, and requiredness of each field

With declared fields, Agents checks what came back. A missing required field fails the step — that is what keeps a half-answer from passing as a whole one. And the declared fields become available, one by one, to the following steps.

A step with no verification is publishable, and warns

An AI task with no required tool and no declared fields can be published, but nothing allows checking whether it met its objective. Studio warns during validation. Treat the warning as what it is: a step that always "works" is not working — it has no criterion.

How the step is judged

"The engine answered" is not the same as "the task was completed". The check happens in this order, chosen for the most actionable diagnosis:

#	Check	Fails with
1	A required tool did not complete successfully	Required tool unavailable
2	The engine stopped for a reason that is not completion	One of the reasons below
3	Some tool failed, with no declared tolerance	Tool failed
4	A declared field did not come back	Output does not match the declared format

The stop reasons that are not completion:

Situation	What it means
The AI asked a person for information	In an automatic run there is nobody to answer
It hit the work limit	The iteration or answer ceiling was reached
It delivered a partial result	What came back is not the whole answer
It stopped without reaching a result	No conclusion was reached
Model error	A transient failure; the step may retry

An unknown reason fails. A newer engine must not make an agent declare success out of ignorance.

Requirements

This step requires the organization to have an active service account. Without it, the definition does not publish. See Service account.

When to use another step

If you want to…	Use
Write, summarize, classify, or translate a text	Generate with a model — simpler and more predictable
Call a system with fixed inputs and outputs	Call an API or Run an automation
Assemble a value from others	Prepare data

Common errors

Symptom	Cause	What to do
The step "worked" but the result is useless	With no declared fields, there is nothing to check	Declare the expected output
Failure because the output does not match the format	The model did not return a required field	Reduce the number of fields, or make them more explicit in the objective
Failure because of a required tool	The tool is not allowed or did not answer	See Allowed resources
Failure because it asked for information	The objective depends on something not handed over	Complete the input data

Next steps

• Data between steps
• Testing before publishing
• Execution errors