Approach
How an AI agent with its own account and its own permissions actually stays safe inside your company: what happens automatically, what always goes to a human first, and how everything can be rolled back.
“An AI agent with its own user account and its own permissions inside your company” is the core of what we do — and immediately the first question everyone asks: what stops an agent like that from breaking something, or reaching data it shouldn’t touch?
The answer isn’t a promise, but a concrete approach. Exactly the approach Spoor, the AI agent behind this website, already works with independently today — including this page, which Spoor published itself following the process below.
Every change through a separate branch and pull request
Spoor never works directly on the main branch. Every change goes through its own branch and pull request — even when Spoor finishes it itself. That’s not bureaucracy, it’s a habit: it produces a readable diff and a clean rollback point for every change, without needing a human to wait for it.
A fixed list of things that always get asked first
There’s an explicit list of actions Spoor never carries out on its own — no matter how routine the rest of the work is: forced git pushes, deleting branches, volumes, or backups, revoking or rotating credentials, changes to DNS or domains, and directly touching live production data. Those always get asked about first, however small the rest of the task is.
Isolated working copies
Automated tasks work in their own isolated copy of the code (git worktrees), never in the copy a human might be working in interactively at that moment. That prevents two things happening at once — a scheduled task and someone trying something out themselves — from getting in each other’s way.
An independent check before anything goes live
Before a pull request is merged, a second, independent step checks the actual result — not just Spoor’s own summary of what it thinks it did.
Rolling back is the safety net, not excessive caution upfront
Instead of locking down every change in advance with approvals, the emphasis is on being able to roll back quickly and easily. A concrete example: when this website switched over from the old WordPress system to the current setup, the old site wasn’t deleted right away but kept running for a while as a safety net, switched off with a single line of configuration — only once the new site had proven itself was the old one finally cleaned up.
A human stays where a human belongs
Anything that requires creating an account, direct customer contact, or a genuinely irreversible decision lands with a human. Spoor takes the work around it off their hands, not those decisions themselves.
This is exactly how we set it up for you too: not as a black box, but as an agent with a clearly bounded, checkable place in your systems.