Skip to main content

​
Assistant Step

An Assistant step runs one of your saved assistants and stores its reply for the rest of the flow.
The step points at a saved assistant; it does not define one. The step holds a reference — subject — to an assistant that exists in your account. Everything the assistant is — its instructions, model, knowledge and integrations — is stored on the assistant itself and is shared by every flow that uses it. Editing those settings from inside a flow edits the assistant, so the change lands on every other flow using it too.The step’s own settings are only the handful listed below.

​
Choosing the assistant

The General tab has a Select Assistant dropdown listing the assistants in your account, and an Edit Assistant Details button that opens the assistant’s own configuration. You can also drag an assistant onto the canvas from the flow editor’s Assistant control, which offers Select Assistant (use an existing one) and Create Assistant (make a new one and reference it).

​
What the step itself holds

​
string
What to ask. Leave it empty and the run’s trigger is used as the prompt — which is what you want for an assistant wired straight to the trigger. When the assistant sits after other steps, map the prompt in:
​
boolean
default:"false"
Include the account’s global instructions in this invocation.
​
boolean
default:"false"
Include the account’s global knowledge in this invocation.
​
string[]
Escalate To — the users who can be assigned as the admin when this assistant escalates a conversation.
The remaining tabs — Provider, Instructions, Knowledge, Integrations, Context — edit the assistant, not the step. They are documented under Assistants.
Two settings that look like step settings are not honoured here. The step always waits for the assistant, for up to five minutes; there is no run-in-background mode on an assistant step, and the Await toggle on the assistant does not change that.

​
What the step returns

The assistant’s response object, stored flat under the step ID. The fields worth knowing:
The answer is at $.STEP_ID.result, not $.STEP_ID.response and not $.STEP_ID on its own.
If the assistant has an output schema, result is an object and you can read into it directly — $.CLASSIFY.result.category. Bear in mind that an output schema is a set of field descriptions given to the model as guidance; it is not validated, so a field can come back missing.

​
Sessions and context

If the run was started with a session — a chat-driven flow, for instance — every assistant step in it invokes into that same session, so a second assistant sees what the first one said. A run with no session gives each assistant step a session of its own, and they share no history. Context variables passed in on the run request are forwarded to the assistant automatically. A step may also declare a context object of its own. The two sets are merged, and the step’s own entries win on a clash — so use one to set a value for a single assistant without changing it for the rest of the run.

​
When an assistant pauses the run

An assistant can stop the flow rather than answer it, and the run parks the same way a Checkpoint does: In both cases the step stores no result and nothing downstream runs. The run is picked up again from the Monitoring page, or by continuing the conversation in chat — the run detail panel offers a Join to Chat button for a paused assistant session.

​
What an assistant step cannot do

Several controls people look for on this step live somewhere else, or nowhere:
  • No output validation or retry loop. An output schema shapes the prompt; nothing checks the reply against it or asks again.
  • No boundaries or guardrail fields. Constraints go in the assistant’s instructions.
  • No per-step memory toggle. Memory is a property of the assistant and its session.
  • No topP, frequencyPenalty, presencePenalty or stop sequences. The assistant’s Provider tab offers Temperature, Max Turns and Max Output Tokens, and that is the whole set.
Editing a flow changes its draft, but editing an assistant from inside a flow saves the assistant immediately — that change is live for every flow and chat using it, published or not.

​
Next Steps

Assistants

Configuring the assistant itself

Condition Step

Branch on what the assistant answered

Checkpoint Step

Hand the run to a person deliberately

Flow Steps Overview

Where step results live