Flows

By default an agent follows a fixed path: it searches your knowledge, then the model answers using its actions. Flows let you draw that path yourself — sort the request first, collect a lead, give a booking link, or hand some topics straight to your team. You build a flow in the dashboard, in the agent's Flow tab.

Simple or Flow#

Every agent has a mode, set at the top of the Configure tab or in the Flow tab's toolbar:

• Simple — the standard path. Nothing to draw.
• Flow — the path you draw. The flow is saved with the draft and goes live when you publish, like any other change.

Switching back to Simple keeps your flow, so you can return to it later.

Whatever the mode, the same protections run before your flow: personal data (card numbers, RIB/IBAN, CIN) is masked, blocked topics are refused, a customer who asks for a person reaches your team, and the turn limit applies.

Steps#

Step	What it does
Start	Where the customer's message arrives.
Knowledge	Searches your knowledge base. The sources it finds can be cited in the reply.
Model	One call to the model with your prompt. The result is stored in a variable; ask for structured JSON output to extract fields such as a name or an email.
Classifier	The model sorts the message into one of your categories — one branch per category, plus Else.
Condition	Picks a branch from rules on variables: equals, contains, greater/less than, is set, is empty, matches a pattern. The first matching rule wins; otherwise Else.
Action	Runs one of the agent's actions (HubSpot, Odoo, calendar link, ticket, HTTP), with arguments you fill from variables.
Variables	Gives variables a value.
Loop	Goes back to an earlier step, at most 5 times, then continues.
Agent	The model answers with the tools you allow, exactly as in Simple mode. Ends the turn.
Reply	Sends a message you write. Ends the turn.
Human handoff	Hands the conversation to your team. Ends the turn.

Messages and prompts can include variables: {{input}} is the customer's message, {{language}} the reply language, and every step that stores a result adds its own (for example {{answer}}, {{knowledge.sources}}, {{lead.email}}). Variables are inserted as text; nothing is ever run as code.

There is no step that runs your own code or connects to an arbitrary server: every call to an outside system goes through an action you configured on the Actions tab.

Templates#

An empty flow offers four starting points, all editable:

• Customer support — sorts the request, answers from your knowledge, and hands off when the sources are weak or the customer complains.
• Lead qualification — extracts the prospect's name, email and need, and saves the lead in Odoo or HubSpot once the email is known.
• Appointment booking — recognises a booking request and gives your calendar link.
• Multilingual FAQ with escalation — answers in the customer's language only from your sources, and escalates otherwise.

From the current settings builds the flow equivalent to your agent's Simple mode, a good place to start customising.

Editing#

Drag steps from the palette onto the canvas, connect an output to the next step, and edit the selected step in the side panel. Undo/Redo (Ctrl+Z / Ctrl+Shift+Z) and Tidy up (automatic layout) are in the toolbar.

Problems are marked on the steps in red as you edit — a step nothing leads to, a branch not connected, an action not chosen. You can save a draft at any time, even unfinished. Publishing an agent in Flow mode is refused until the flow has no problems, and your test cases then run against the flow as usual.

Testing#

Test saves the draft and sends a message through the real runtime, as the playground does. The reply appears with every step that ran, how long it took and the tokens it used, and the path taken lights up on the canvas.

Limits#

To keep every conversation answered, a turn runs at most 25 steps and 60 seconds, and a loop repeats at most 5 times. Past a limit — or when a step fails, unless it is set to continue — the conversation goes to your team. Model calls in a flow are billed exactly like the rest of the agent's.