This page is machine-translated from French. Read the French original.

Agent flow

A flow is a predefined sequence of steps, visually constructed, and always executed in the same order. Whereas an agent decides what to do, a flow does what was planned.


When to prefer a feed to an agent

Agent Flow
Decide on the steps The model YOU
Reproducibility Variable The same steps are followed each time; the text produced by a block of instructions may vary.
Adapting to the unexpected Good None
Number of calls to the model Pupil Reduced to the steps that need it
Debugging Difficult Step by step

Consider a flow where the procedure is stable and the result must be constant: a daily extraction, a standardized processing, an API call followed by formatting.

Hire an agent when the approach depends on what you find along the way.


The manufacturer

Settings → Agent → Agent Skills, Agent Flow section, then «Create a flow». For administrators only.

The available blocks

Startup. Declares the input variables of the stream — what is provided to it at startup. A variable is mandatory, optional, or fixed; a fixed variable remains invisible to the model.

API call. Sends an HTTP request: URL, method, headers, request body. The response is stored in a variable. An option allows the response to be sent directly to the user, bypassing the model—useful when the result is already usable. This block is not subject to any approval: the request is sent without human validation, regardless of the method used.

Instructions for the model. Give the model an instruction, applied to the available variables: summarize, extract, reformat, classify. The result is stored in a variable.

Web extraction. Retrieves the content of a URL and stores it in a variable. The same direct return option is available.

The blocks are chained together as many times as necessary. As soon as a block fails, the flow stops and returns the error.


An example

Objective: each morning, to collect new publications from a regulatory source and produce an internal memo.

  • Start — variable source_url.
  • Web extraction — retrieves content from source_url → contenu_brut.
  • Instructions — «Extract only publications dated yesterday or today. If none, answer exactly: NONE.» → nouveautes.
  • Instructions — «Write an internal memo of no more than ten lines based on nouveautes, citing the references. » → note.
  • API call — publishes note in the internal distribution tool.

This workflow, associated with a scheduled task, runs every morning without intervention (in single-user mode, the only mode where scheduled tasks exist). It then runs without any human reviewing the note: a workflow has no conditions, and the API call will execute even if the instruction returned NONE.


Use a stream

A saved and active feed becomes a skill usable by agents throughout the instance. It is active upon creation and can be deactivated from the Agent Feed section. A feed can therefore be called:

  • from a conversation, by entrusting the task to the agent, who decides to launch it and provides him with the variables;
  • from a scheduled task.

One stream cannot call another.


Best practices

Test block by block. The builder doesn't have a run button: add a step, save, launch the flow from a conversation with @agent and check its output in the flow, rather than writing the whole thing before the first attempt.

Name your variables explicitly. liste_factures_impayees better than resultat2 — the readability of the feed depends on it, and you will come back to it in six months.

Plan for the empty case. What should happen if the API returns nothing, if the page is inaccessible, if there is no new content? An explicit instruction on this case avoids absurd outcomes.

Constrain the format of the model's outputs. «Answer only with a bulleted list, without introduction» produces a result usable by the next step; a vague instruction produces a text to be reworked.

Beware of secrets. An API call may require a key. This key is visible to administrators in the flow definition, and the request body is displayed in the conversation flow: use a dedicated key with limited privileges. Treat these flows with the same care as a production script.