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

Agent skills

A skill is a tool that an agent can call. They are activated at two levels: globally by an administrator, then space by space.

Where to activate them

Réglages → Agents (administrator) — Enables or disables a skill for the entire instance, and configures those that require it (access to a database, email, a search engine). A skill disabled here is not available anywhere.

Space settings → Agent — chooses, from among the authorized skills, those that the agent of this space can actually use, and sets its model.

This dual level allows a sensitive skill to be authorized for the instance while only activating it in a single space.

Catalog

Search and web

Web search — queries a search engine and reports the results. Requires a search engine to be configured; several services are supported, including self-hosted engines for installations without external dependencies.

Reading a web page — retrieves the content of a URL and extracts the relevant text. Works without configuration.

These two capabilities send requests externally. On an isolated installation, they are irrelevant.

Documents and files

Document summary — summarizes a document from space by relying on its full text rather than extracts.

File system — lists, reads and writes to an authorized directory. Each elementary operation can be disabled separately — for example, allowing reading and prohibiting writing.

File creation — produces documents (text, spreadsheets, office documents) that the user retrieves at the end of the task.

Data

SQL Agent — connects to a database to list tables, read their structure, and execute queries. Supported database engines include PostgreSQL, MySQL, and SQL Server.

Create a dedicated read-only account. The agent builds its own queries. Limited access to only the necessary tables is best practice, and the only real protection.

Graphs — produces a graph from data collected during the task.

Communication

Gmail And Outlook — search, read and prepare messages. The design prioritizes drafts submitted for human validation rather than direct sending.

Google Calendar — checks availability, reads and creates events.

Microsoft Teams — reads discussion threads and posts messages to them.

These connections are configured once by an administrator. On a multi-organization instance, each organization connects its own access.

Memory and knowledge

Conversation history — research into previous exchanges of space.

Second Brain — Consult and enrich your personal knowledge base. → Second Brain

Skills library — Loads a job instruction sheet upon request. → Skills library

Creation

Image generation And video generation — produce a visual during a task. Require a configured provider. → Images, videos and voices

Meetings

Meeting minutes — structure a transcript into a report: participants, decisions, actions to be taken.

Interaction

Request for clarification — allows the agent to ask you a question instead of making assumptions when the instructions are ambiguous. They wait for your answer before proceeding.

Job skills

SEO Audit And Prospecting — two complete business applications, accessible both via conversation and through their own interface. → Business agents

Skills active by default

A small set is activated from the outset, because it is useful everywhere and without risk: conversation history, document summary, web page reading, and the two embedded business skills.

Everything else is activated on demand.

Intelligent skills selection

The more tools an agent has, the more space the description of those tools takes up in its context — to the detriment of the task itself.

WivenLLM can therefore only present the skills to the agent relevant to the current application, rather than the whole set. The agent retains access to the set; he simply doesn't see them all on each turn.

The benefit is twofold: less context consumed, and better tool choices — an agent who sees five relevant tools makes fewer mistakes than an agent who sees fifty.

Choose what to activate

Principle: the bare minimum. Each additional skill expands what an agent can do — including by mistake.

Questions to ask yourself before activating:

• Who has access to this space?
• What can happen if the agent misuses this skill?
• Does this capability send data externally?
• Is there a human validation process before the final effect?