Español
English
Català
GitHub
Designing roles and permissions for AI agents, step by step
An agent should be able to do exactly what it needs and nothing more. Here's how to design a role for a follow-up, sorting or customer service agent.
The principle is the same as with people: everyone gets access to what they need for their job. The difference is that an agent acts much faster, so one permission too many has bigger consequences.
1. Describe the task in one sentence
"Remind clients about unanswered proposals." If it doesn't fit in one sentence, it's probably two agents.
2. List the minimum actions
- View clients and opportunities.
- Create tasks.
- Send messages using an approved template.
3. Explicitly deny anything sensitive
Changing amounts, deleting data or changing permissions should never be part of an autonomous agent's role. If it tries, the action is blocked and logged.
Blocks in the log aren't errors: they're proof that the limits work.
4. Review after 30 days
With a month of logs, you'll know which permissions the agent really uses and which you can remove.