Skip to content
ClaudeClaude

How Lovable vibecodes production software at scale

Vibecoding a beautiful, working prototype is easier than ever. Running a platform that non-developers use to ship production software is a different engineering problem. Today, software built on Lovable serves 600M+ monthly sessions. In this session, Fabian Hedin, Cofounder and CTO of Lovable, walks through the systems Lovable built to run Claude reliably at consumer scale — the fleet-learning layer that catches coding mistakes, the eval loop that gates every model release, and how to keep Lovable itself improving.

Fabian Hedinhost
May 20, 202631mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

Lovable’s self-healing platform reduces user “stuck” moments at scale

  1. Lovable evolved from the early GPT-engineer demo into a chat-plus-preview product used to build everything from landing pages to complex web apps at massive scale.
  2. The core product challenge is that users—especially non-technical ones—hit “stuck” states where iterative prompting fails, making completion and publishing much less likely.
  3. Lovable defines and detects stuckness (e.g., repeating the same request, complaining, or abandoning) and treats stuck causes as three buckets: prompt-solvable, platform-edge gaps, and larger platform investments.
  4. “Lovable Overflow” is a continuously pruned issue→solution corpus that retrieves relevant prior fixes and injects adapted context into the agent to skip unproductive back-and-forth.
  5. A “venting” tool lets the agent file structured frustration reports that are deduped, investigated, and turned into PRs—shipping ~10 fixes/day and even surfacing production incidents early.

IDEAS WORTH REMEMBERING

5 ideas

Non-technical UX fails when users become “hard stuck,” so preventing stuckness is a core product feature.

Lovable treats stuck moments as the worst-case experience because non-technical builders can’t easily drop into code or tooling to diagnose problems, so the platform must proactively reduce friction before users spiral into repeated failed attempts.

Operationalizing “stuck” as a measurable metric enables systematic improvement.

Lovable’s is_stuck heuristic (repeat requests, complaints, or abandonment) plus a small classifier turns a subjective frustration state into data that can drive retrieval, tooling fixes, and platform roadmap decisions.

A curated “issue → solution” corpus can outperform naive retries by skipping to proven fixes.

Lovable Overflow searches prior problem/solution descriptions (from broad best practices to package-version specifics) and injects tailored context into the main agent, reducing back-and-forth, latency, and cost.

Knowledge must be continuously pruned because it decays and can harm outcomes.

They track a success ratio per knowledge artifact and withhold or remove items that become stale (e.g., library updates), emphasizing that retrieval systems need deprecation mechanics as much as ingestion mechanics.

Letting the agent report tool/platform pain creates a fast self-healing engineering loop.

The vent tool sends structured feedback to Slack; another agent dedupes and investigates, then proposes PRs for human review—yielding rapid fixes such as handling filenames with spaces and other “small but blocking” issues.

WORDS WORTH SAVING

5 quotes

We wanted to create something for people who didn't necessarily know how to code, um, and did not have these capabilities at all really to create, uh, software. Um, and, and we call that the 99%, right? Uh, uh, not the 1% that can code, but, uh, the rest of the population.

Fabian Hedin

It talks about that the last 10% of code takes 90% of the time, and the other part of this quote is that, uh, the first, uh, 90% also takes 90% of the time. So you end up with 180% of the time.

Fabian Hedin

So really our vision with Lovable, um, on, on the technical side is that, uh, every app that is built on the platform should help improve the next.

Fabian Hedin

So we have this metric internally that we call is_stuck, and, uh, it will be true if you're asking for the same thing three times in a row. So if you're ask- asking, "Fix it, fix it, fix it," we will assume you're stuck.

Fabian Hedin

Several times now this, uh, this Slack channel with the agent venting has been kind of the first signal for us to identify a production incident.

Fabian Hedin

Chat-to-app abstraction (preview-driven development)Production-grade software for non-technical usersDefinition and classification of “is_stuck”Three categories of stucknessLovable Overflow (retrieval + context injection)Knowledge pruning, staleness, and success-ratio trackingAgent “venting” to Slack → dedupe → PR pipelineIncident detection via vent spikesMetrics: stuck rate, publish rate, PR merge throughputModel changes and retuning self-healing systems

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.