Blog/Product
Act, don’t ask: how Axer’s agent works
It plans, it edits real files, and it only stops when the next step should be yours.
Axer Team · 15 November 2026
Most coding chat windows wait for you to say “go on.” Axer’s agent is built the other way around. You describe the work. It makes a short plan, starts editing the repo, and keeps going until it hits something that should not happen in silence.
A plan, then the files
The agent reads the project first. It names the files it will touch and writes a brief plan — enough to follow, not a essay. Then it continues on its own. You do not have to prompt it after every file.
In the chat you see live step rows: the path it is in, a small +/− count, not a wall of patches. Click a row and the file opens in the editor.
It asks before it acts
Preview links, publishes, and shell commands wait on an Accept button. Nothing lands on a live host, and no command runs, until you say so. If you accept a kind of step once in a project — a preview, a publish, a safe command — you can tell it to remember. After that it will not ask again for that kind of work, in that project. You can reset this in project settings.
Risky commands are still asked every time. Auto mode, which picks a model for you, never shows the routed model name. You just see Auto.
Checkpoints, always
Before each turn, Axer takes a snapshot of the project files. If a change is wrong, you restore that point. The current state is saved first, so a restore can itself be undone. There is a longer note on this in Checkpoints.
What you see
The workspace is one product: file tree, editor, preview, and chat. The agent can open the preview in the built-in browser, read console and network errors, and fix them. When you are ready to ship, it publishes — and if the target is a live cPanel or Plesk host, it takes a backup first.
That is the contract. You set the work. It does the work. It stops when the next decision should be yours.