Is it safe to give an AI agent access to our code repositories?

How it works

In Lua, a person sets what each agent may do with each tool. Governance rules block or pause named tool calls and scan every incoming message for prompt injection. Every step and tool call is logged, and the log can be sent to your own monitoring tools. Roles for viewers, operators, builders, admins and owners decide who can change any of it.

What a person keeps

Engineers keep the merge. The fleet gets read access first, then the right to open pull requests. Nothing deploys without a person.

What to watch for

Ask for the audit trail, not the certificate alone. A certificate says the vendor has controls; the trail shows you what the agent did in your repository last night.

An example

The Security Engineer agent upgrades five vulnerable packages and marks the pull request merge ready. The Engineering Manager agent routes it to Tom, because dependency upgrades need an engineer. Tom approves, and the merge is in the log.

Example data is invented unless a customer is named.