AI EngineerJune 2, 202619m

How Lovable self-improves every hour — Benjamin Verbeek, Lovable

TL;DR

Lovable is learning from failures every hour: Benjamin van Beek says the goal is simple, a mistake should happen once and then never again, using production signals from 200,000-plus projects created per day.
They built a Stack Overflow for agent mistakes: when a user is clearly stuck, then later gets unstuck, Lovable extracts the successful fix, clusters similar cases, and injects that context into future runs only when relevant.
They validate knowledge in production with holdouts: a lightweight model sometimes injects a known solution and sometimes injects nothing, then Lovable compares which projects succeed more often so stale or harmful context gets removed.
The agent can vent directly to engineers: Lovable added a "send feedback" tool so the model can complain about missing tools, bad docs, broken platform behavior, or repeated failures caused by the environment.
Agent complaints exposed bugs humans missed: within an hour of launch, about 20 reports pointed to file copy failures on filenames with spaces, then non-breaking spaces from WhatsApp and Mac screenshots, which led to a proper fix.
Complaint spikes became an incident detector: when sandboxes or other platform pieces broke, vent volume spiked immediately, giving the team a fast, specific signal about what users and agents were running into.

The Breakdown

Lovable is turning user frustration into a live self-improvement loop: it mines stuck sessions into a Stack Overflow-like knowledge base, and even lets the agent complain directly to Slack when the product or tooling gets in its way. Benjamin van Beek says that system is already reducing repeated “fixing intent” messages, increasing deployments, and catching incidents fast enough that spikes in agent complaints can signal outages.