Lovable rolls out Agent Mode in beta

Most AI coding tools work like a conversation. You describe what you want in a prompt, the tool answers, and you carry the result back into your project yourself. Lovable presented Agent Mode as a departure from that pattern: the company described the mode as enabling Lovable to think, plan and take actions autonomously rather than only answering prompts ^[1].

What the mode does

Inside a project, Agent Mode searches the codebase to locate the exact files, functions or components it needs ^[1]. Rather than loading a whole project at once, it reads files on demand and applies edits with full project context ^[1]. That distinction matters in practice: a project contains far more code than fits usefully into a working context at any one time, so finding the few relevant places first and reading them selectively keeps the edit anchored to what the rest of the project actually expects.

The mode also takes over work that users previously did by hand. It inspects logs and network activity to debug errors without the user pasting them in ^[1]. Log lines and network requests are where runtime problems usually show themselves, and copying the right fragment out of them is a step people often get wrong; here the tool looks for itself.

Beyond the project, it can search the web in real time for documentation, content, images or screenshots, and it can generate or edit images for the app ^[1]. So the material the agent works with is not limited to what already sits in the repository: reference material and visual assets can be fetched or produced as part of the same run.

How it was released

Lovable began rolling out Agent Mode in beta from 30 June 2025 ^[1]. Access was gradual rather than universal, and the mode had to be switched on in a project's settings ^[1] — so whether a given project could use it depended on that switch as much as on the rollout reaching the account.

Why it matters

The change is in the unit of work. Answering a prompt produces text that a person must then act on; planning and acting produce a sequence of steps taken against the project itself — locating code, reading it, editing it, checking logs, looking things up. The described capabilities are all about giving the tool the means to gather what it needs on its own instead of asking the user to supply it. Being a beta, with access gated by a per-project setting, it was an early form of that way of working rather than a settled one.