Why AI agents should propose and people should approve

A suggestion is not an action

That is useful, and it is also very different from changing your product. In SylvieSI, agents have no tool for sending a message, opening an issue or releasing a change.

Those actions belong to controlled services. A service carries out an action only when product policy allows it and, where the policy requires it, a person has approved it.

What an approval covers

An approval names one exact change: what it is, where it applies, which environment, who it reaches, how risky it is and how it can be undone. If anyone edits the change afterward, the approval stops applying and the change returns for review.

For high-risk actions, the person who asked for the change cannot approve it alone, and the approver types a confirmation. A workspace with a single member is the exception, and the self-approval is logged as such.

Nothing merges on its own

When SylvieSI prepares a fix, it opens a draft pull request on a repository you approved. Your engineers review it and merge it. Deployment is a human decision too.

Six boundaries never change: no automatic merge; no silent risky deploy; no arbitrary generated JavaScript on your pages; no unrestricted proxy to your systems; no autonomous change to a database schema, payments or authorization; and no claim of an external effect before the provider confirms it.

What agents do instead

Ten agents each take one job, from triage to the engineering handoff. They do the reading, the testing and the drafting. Each run is saved with its sources, its tools and its cost, and output that cites a source that does not exist is rejected. That leaves the decisions to your team.

What we have not done

SylvieSI holds no security certifications, and no external security audit has been carried out. The Trust and control page lists these limits, and we update it when they change.