On November 12, Alan, AWS and Qovery filled a room in Paris to talk about running AI agents in production. Almost nobody raised a hand when asked who gives an agent full access to production. The demo that night was a video. This time we run it live.
Most teams keep a human on every change the agent makes. It feels safe, but it isn't. Anthropic published that 97% of Claude Code permission prompts get approved. After a few hundred prompts, people stop reading and click yes. You pay for the review with speed and you don't get the safety.
The real problem is permissions. Roles were made for humans, and a role that can change something can usually delete it too. Fix that, and an agent with bad judgment can only do narrow damage.
We'll show how to give each agent exactly the access its job needs, checked by the API on every request, whether the call comes from the console, the CLI, Terraform or MCP. Then we hand the agent production and watch what it is allowed to do, and what it is not.
What you'll learn
- Why human approval is not a safety net: what the 97% approval rate means for your review process
- Why agents still wait for a human: judgment, reach, and permissions that are not precise enough
- Permissions written for agents: one policy per agent, deny by default, checked on every API call
- Live: an agent on production: Claude Code connected to Qovery through MCP. It investigates, deploys a preview and cleans up, then tries to delete the production database and gets blocked
- Live: the same block through Terraform: the agent runs
terraform applyand hits the same rule, because every interface goes through the same API - Live: writing a policy in plain English: from "let my agent resize my production apps and nothing else" to a tested policy and a real token
- Which jobs to hand over first: cleanup and cost, then tests on a copy of production, then security patches
Speakers
Guillaume Da Silva - Staff Solutions Architect at Qovery
Romaric Philogène - CEO & Co-founder at Qovery
Details
- Date: December 1, 2026 at 5pm CET / 11am EST / 8am PST
- Duration: 60 minutes including live Q&A
- Format: Talk plus live demos, with Q&A at the end
This session is for platform teams, SREs and engineering leads who already let agents read production and want to let them change it without a human on every click.
