This page is machine-translated from French. Read the French original.
The agents
An agent is an assistant authorized to act, not just respond. Where an ordinary conversation produces text, an agent performs a series of concrete actions: searching the web, reading a file, querying a database, preparing an email, producing a graph.
Trigger an agent
In a conversation
Prefix your message with @agent :
@agent Cherche les dernières évolutions réglementaires sur la facturation
électronique en Suisse, puis résume-les en dix lignes.
The agent takes control, answers, and then waits for your response for 5 minutes: as long as you respond, the session continues. After this time, or after the order is placed, the session will stop. /exit, The session ends and normal conversation resumes.
Permanently
A workspace set up in mode Agent assigns each message to an agent, provided that the model and provider support native tool calls; otherwise, it is necessary to go through @agent. Reserved for spaces dedicated to automation: this mode is slower and less predictable.
Without human intervention
A scheduled task runs an agent according to a schedule, without anyone being in front of the screen. This function only exists in single-user mode.
How an agent works
- He reads the request and determines what he needs to do.
- He chooses a tool from among the skills activated on the instance.
- He executes it and observes the result.
- He starts again — as many times as necessary, within the limits below, each result guiding the next step.
- He responds, based on what he has actually obtained.
The sequence of events is visible in the conversation: you can see which tools were called and what they returned.
An agent is supposed to say that it has failed rather than invent a result, but this is not guaranteed: over a long series of actions, a model may announce as done a step that it has not executed.
Request one or two actions at a time, and rely on the flow, approval cards and the actual result (the email inbox, the file) rather than the agent's summary.
Limits are put in place to prevent a stuck agent from spinning indefinitely:
| Limit | Default value | At the deadline |
|---|---|---|
| Tool calls by response | 10, configurable by an administrator | The last tool runs, then the agent writes its response |
| Rounds of a session | 100 | The session ends |
| Awaiting approval | 2 minutes | The action is denied |
| Awaiting your response | 5 minutes | The session ends, and the interface signals this. |
An agent session remains within the model's capacity: previous tool results are shortened, and the oldest exchanges are summarized. A bar below the input area indicates the context is being filled.
What determines the quality of an agent
The model
This is the main factor. Using tools is more demanding than conversing: you have to choose the right tool, build valid arguments, and interpret a sometimes unexpected result. A model that answers questions very well can fail as an agent.
Each space can therefore assign the agent a different model than the one used for conversations. If agents matter to you, give them the most capable model you have available.
The formulation of the request
An agent follows instructions. Vague instructions produce vague results.
Ineffective: «"Take care of competitive intelligence."»
Effective : «Consult the Pricing pages of the following three websites, note the prices displayed for the entry-level offer, and present them in a comparative table with the date of consultation.»
Available skills
An agent can only do what its activated skills allow. A request that assumes a missing tool fails — and the agent must report it.
What an agent can do
Depending on the skills activated:
| Domain | Examples |
|---|---|
| Web | Search, view a page, extract its content |
| Documents | Search the documents in the space, summarize one, produce a file |
| Files | List, read, and write to an authorized directory |
| Data | Query an SQL database, produce a graph |
| Communication | Read, prepare and send an email, view and edit a calendar, search for a colleague in the Microsoft directory and check their Teams presence |
| Memory | Search through your past conversations, leave a note in the Second Brain |
| Creation | Generate an image or video |
| Job | Auditing a website, researching prospects (workstation application) |
What to watch out for
An agent is truly taking action. A file writing skill actually writes files; an email skill can send real messages. Skills are activated for the entire instance, not space by space: only activate what you need.
Human approval does not cover all actions. Each skill decides for itself whether it requires validation; there is no central control. An approval card appears in the conversation for certain actions only:
| Actions | Approval |
|---|---|
| Send an email or a draft, reply, delete a draft (Outlook, Gmail); move an email to the trash (Gmail) | Every time. «Always allow» is not an option. It's denied when no one can respond, for example, in a scheduled task. |
| Write, edit, move, or copy a file; create a document; create or edit a draft; open email attachments; sort an email (Gmail); create or edit a calendar event; create a scheduled task | Approval card, but «Always allow» remains possible, and a scheduled task approves them automatically. |
| SQL queries, MCP server tools, agent flows (including their API calls), web search and reading, Teams tickets, Second Brain | No approval required: the action is executed as soon as the agent decides it. |
An administrator can also approve certain installation skills in advance, except for those in the first line.
An email can contain malicious instructions. An agent reads messages written by unknown senders. Read the approval card before validating a message: it indicates the recipients and the subject, and, for Outlook, the beginning of the message and external recipients.
An agent can be confidently wrong. It reports what its tools returned, without verifying it. An erroneous web page remains erroneous after being read by an agent.
An agent costs more than a response. Each step is a call to the model. With an external provider, a ten-step task costs approximately ten times that of a simple question.
One agent is slower. Allow anywhere from tens of seconds to several minutes depending on the task, model and equipment.
Expand the agents
| Need | Mechanism |
|---|---|
| Add business instructions loaded on demand | Skills library |
| Connect a standardized external tool | MCP Servers |
| Constructing a deterministic sequence, without a code | Agent flow |
| Run automatically, according to a schedule | Planned tasks |
| Use a complete business application | Business agents |
