AI Agents
Use an Agent when a step needs interpretation, such as classifying a request or extracting a structured summary. Configure its instructions, supported model, input fields, output format and budget. Select the data to send through explicit mappings; an Agent does not automatically receive the entire triggering record.
Include values in instructionsโ
Insert values from the trigger or an earlier step into your instructions, such as
Summarise {{record.attributes.name}}. The model receives the captured value when the step runs.
Optional missing values, nulls and skipped steps become empty text. Missing required outputs fail
the run. Supplied values are task data and cannot change the Agent's instructions or capabilities.
Publication checks that the selected fields and earlier steps exist. Preview retains the configured
references because it does not execute steps or call a model.
See the reference syntax.
Choose the outputโ
| Format | Result available to later steps |
|---|---|
| Text | output.text, a text result. |
| Classification | output.option_id, one configured option's UUID. Publication requires 2โ50 options. |
| Structured | The fields of your configured object schema. |
Keep classification option identities stable when renaming their labels. For example, define Urgent and Routine, then use If/Else to compare the accepted option ID with the Urgent option's ID. Compare identities, not the label or an arbitrary model sentence.
Outputs are validated before acceptance. An invalid result fails the step even if the model consumed tokens. A later ordinary Task or record action can perform work based on an accepted result.
Select capabilities deliberatelyโ
Agents start without tools. Supported capabilities include permitted Task and record reads, bounded web search/fetch, and optional human clarification. Local Task/record tools are read-only; configuring an Agent does not give it general write access.
For permitted record reads, the Agent receives the object IDs and metadata for its selected readable fields: slugs, labels, descriptions, choice IDs and labels, numeric bounds, currency codes and relationship targets. Map a record ID into its input; object IDs and choice dictionaries do not need to be copied into the instructions. Record values still require an explicit permitted read.
Describe the meaning of numeric scales in the field definition, for example โ1 low to 5 highโ; bounds alone do not define their meaning. Metadata is treated as untrusted data and cannot widen access. It must match the published schema fingerprint and current permissions. Republish after schema changes. Schema context is limited to 32 KiB; reduce the selected fields if it exceeds that limit.
Configure any permitted web domains and request limits, and inspect the returned sources. An activity entry proposing a tool is not proof that it executed. Accepted final citations remain with the result even if conversational activity later expires.
Human clarificationโ
If enabled on the Agent step, clarification permits single-choice questions, free text, or both. Set the maximum number of questions, response timeout and text length; choose the assignee on the workflow's Agent step.
The Agent creates a question Task and waits. An answer returns to the same Agent execution. Expiry returns an unavailable-answer result. Cancellation closes pending questions without resuming the model. Human waiting does not consume active execution time, but already consumed calls, tokens and cost are not reset. See Human questions.
Budgets and versionsโ
Budgets bound total tokens, model calls, local read calls, active duration and estimated provider cost. Web and question limits are separate, while their model rounds still use the cumulative model budget. Leave enough model calls for the Agent to process clarification answers.
Provider-cost estimates are not customer invoices or a guaranteed invoice cap. Pending or unknown usage is not zero. Inspect the run's confirmed usage and any uncertainty after a failure.
Publish the workflow to freeze its Agent instructions, output format, capabilities and budgets together. Later edits belong to a new workflow draft and take effect when that workflow version is published. Existing runs retain their original configuration. Preview does not call the model.
The Agent API reference describes output schemas, available capabilities and budget fields.