Lovable details the engineering rebuilds behind 70 million projects

From GPT Engineer to Lovable

In the spring of 2023, Lovable’s founder released an open-source project called GPT Engineer on GitHub. Users described the software they wanted, answered clarifying questions and received a complete codebase.

Later that year, the plan for Lovable was to let people who could not code create, deploy and improve software through chat while seeing the result instantly. More than 100,000 pull requests later, the company says its team, including co-founder Fabian, has turned that plan into a platform where users have created more than 70 million projects.

Lovable initially focused on new applications rather than existing codebases. It also chose an opinionated technology stack, giving the company more control over libraries, code structure and the execution environment. That made it possible to inspect failures and improve the system across projects using the same setup.

Why Lovable reduced agent autonomy

Lovable initially explored more autonomous agents. An engineering design allowed an agent to create tasks and delegate work to subagents, but the company found that less constrained agents could become trapped in technical problems, make errors and take actions without a predictable finish time.

The company responded by removing almost all of that agent code and returning to a more predictable sequence of tools. By late 2024, the main editing flow used a single response containing requested file changes, with chat and a live preview. At the time, native tool calling was not reliable enough for Lovable, so the model wrapped code in XML-like tags such as <lov-write>, which the system parsed after the response.

Lovable also used fine-tuned small language models to identify the code relevant to each requested change. This helped the main coding model receive the necessary context, including files that did not need editing.

Knowledge files and faster feedback

Instead of immediately training models on failures, Lovable’s engineers reviewed chat logs from broken applications and wrote short notes about recurring mistakes. These “knowledge files” were retrieved when relevant and added to the model’s context.

One example instructed the system to use only icons available in lucide-react and to choose a generic icon when uncertain. Lovable later applied the same approach to Lovable Stack Overflow, which retrieved problems and solutions for the agent. The company says an early version reduced the rate at which users got stuck by 5%.

To shorten the edit, check and correction cycle, Lovable built sandboxes that retained project files and kept development servers running. Incremental TypeScript checks reused earlier work, allowing errors to be returned to the agent while the project continued running. The live preview could update without creating a fresh production bundle.

Forking entire working environments

Lovable later extended its infrastructure beyond saving source files. Its stateful sandboxes checkpoint a machine’s memory, CPU state and disk, then restore them when needed. The company says its Camille framework can fork and merge entire running computers.

This allows Lovable to fork an application together with its working environment and database, giving agents a way to test different approaches from the same starting point. The company presents this as a foundation for taking on more complex software work.

Lovable says three major problems remain open: automating testing, deployment and maintenance; making all types of software seamless across different environments; and building enough trust to prevent misunderstood requests from deleting data or causing irreversible damage. These challenges define the next stage of the platform’s development.