Advanced Behavior
Human-in-the-loop questions inside a chat, embedded chat interfaces, and the current limits of workflow chats.
Coming soon
Human-in-the-loop inside a chat
A workflow turn doesn't have to be a single request and response. A workflow can pause mid-run to ask the user a question and resume once they answer — all inside the same conversation.
When the workflow suspends with a question, the chat renders the prompt and waits. The user's answer is sent back to the running execution, which resumes from where it paused. This is the same suspend/resume mechanism described in Human in the Loop, surfaced conversationally: instead of an approval inbox or a Slack button, the question appears in the chat thread.
Use it for flows that need a clarification before continuing — confirming an action, collecting a missing field, or choosing between options — without leaving the chat.
Embedded chat
Setting the trigger's Mode to Embedded Chat keeps the workflow out of the hosted chat
listings — the AI Hub composer's Workflows cascade and the Community Edition Chats
page — and lets you drive it from your own UI. Your interface calls the
New Chat Request webhook with the
same { message, conversationId, attachments } body and renders the streamed reply however
you like. The workflow contract is identical; only the front end changes.
This is the path to embedding a branded assistant in your own product while still building the logic as an ordinary ByteChef workflow.
Limitations
Keep these in mind when designing a chat workflow:
- The chat trigger runs synchronously. A turn occupies the chat until it returns; design workflows so a turn completes in a reasonable time, and lean on streamed output for long answers.
- Access follows the workflow. A user who can see a workflow's chat can run it; scope chat workflows the same way you scope any other workflow.
- Conversations are environment-isolated. A conversation is bound to the environment and workflow version it started against.
- A workflow chat is not the assistant. It runs your workflow, not the AI Hub agent, so the assistant's tools, memories, and specialist subagents don't apply inside it.
Keep turns responsive
If a chat feels slow, stream partial output early in the workflow and defer heavy work, so the user sees progress instead of a spinner. See Streamed replies.
How is this guide?
Last updated on