English pilot — the directory is mainly in French for now.
AI agents are useful precisely because they act on their own — and that is also what makes them risky. Here are the four risks users actually run into, and the guardrails that work, without alarmism or naivety.
An agent almost never says “I don't know”: it produces fluent answers, including when they are wrong — an invented price, a shifted date, a summary that flips a nuance. The danger is not the glaring error, which you spot, but the plausible one you let through. Guardrail: for anything that commits you (money, customers, reputation), require cited sources and reread the passages that matter — or set up a human approval before the action.
A pay-per-use agent that locks onto the wrong target can repeat the same action hundreds of times: documented cases show coding agents spending several hundred dollars of API calls in a single badly framed workday. Guardrails: a maximum number of steps per task, a spending cap per day and per agent — with a real stop, not just an alert — and a cost report you check every week. For technical uses, LiteLLM is the tool designed exactly for this; Langfuse makes every run replayable after the fact.
To do its work, an agent reads your emails, your documents, sometimes your entire folders. That data flows through the tool's publisher and the model's — often in the United States. Guardrails: authorize only the folders that are needed, check where the data is hosted and whether it is used to train the models (and opt out when you can), and for truly sensitive data, prefer models that run locally (Ollama (local agents)) or a European publisher such as Mistral Vibe (formerly Le Chat). In Québec, Law 25 governs the communication of personal information outside Québec; in Europe, it's the GDPR: in both cases, the responsibility remains yours.
Most incidents recounted by early users have the same cause: an agent allowed, from day one, to send, pay or delete. Guardrails: start read-only, then in “draft” mode (the agent prepares, you send), and only open autonomous action on tasks whose results you have reviewed for several weeks. Ban by default: payments without approval, sending to new recipients, permanent deletion.
Treat an agent like a very fast employee on probation: a written scope, limits in numbers, a log of what it has done, and responsibilities that grow with demonstrated trust — never the other way around. Every profile in this directory ends with the guardrails specific to the tool; take them seriously, they come from real incidents of the first users.
Browse the directory's 111 agent profiles →