From the source
Lovable's co-founder explains how the company repeatedly rebuilt its AI app builder, narrowing the problem and constraining the agent to keep it dependable.
- The team chose greenfield projects and an opinionated stack, giving it control over libraries, code structure and the environment the generated code ran in.
- An async agent design that let the agent create tasks and delegate to subagents got stuck in technical rabbit holes, so the team deleted almost all agent code and returned to one request producing a predictable chain of tools.
- By late 2024 the main editing flow was a single response of file changes, with the model wrapping code in XML-like tags such as <lov-write> because native tool calling was not reliable enough.
- Instead of the fine-tuning the 2023 plan called for, engineers read chat logs from broken apps and wrote short knowledge files, which a retrieval step added to the model's context.
- The company says that work was critical to its late-2024 take-off, and an early version of Lovable Stack Overflow cut the rate at which users got stuck by 5%.
