Lovable, an AI app builder for non-technical users, has introduced a novel approach to continuous product improvement: letting its AI agent complain about its working conditions. The platform created what it calls a "vent tool" that captures moments when the agent encounters friction, enabling the team to identify and fix problems before users experience them. Around 200,000 projects are created on Lovable every day, generating a wealth of data about where users and agents get stuck.

Why Do Non-Coders Get Stuck Differently Than Engineers?

The challenge Lovable faces is fundamentally different from other AI coding tools. When a software engineer hits a roadblock while using an AI assistant, they can usually jump in, fix the issue manually, or guide the agent toward a solution. Non-coders don't have that safety net. If they get stuck early in a project, they're often far into unfamiliar territory and debugging becomes nearly impossible. The data tells a stark story: users who get stuck early are 4 times more likely to leave the platform entirely.

This creates a different success metric for Lovable than for traditional coding agents. While other tools might tolerate occasional friction, Lovable must minimize stuck moments to nearly zero. Even a single stuck moment can mean losing a user who might otherwise have completed their project.

How Does Lovable Detect and Solve Agent Frustration?

Lovable's team built two automation loops to continuously improve the platform and move toward a vision where any issue is experienced only once. The first involves detecting when users are stuck using what the company calls LLM (large language model) judges, which are external reviewers trained to identify when a project isn't making meaningful progress. These judges look for telltale signs like repeated requests for the same thing, complaints about implementation, or users giving up on sessions they would otherwise continue.

The team categorizes stuck moments into three types:

• Stuck but solvable: A tricky bug that requires hard prompting and iteration, but once the core issue is found, the agent can resolve it
• Stuck due to tool limitations: A configuration that needs changing but the agent lacks permissions, or a simple fix that's trivial in principle but impossible with current tools
• Stuck due to fundamental constraints: Something the tech stack doesn't support, which would require significant engineering effort to add

To solve the first category, Lovable built what it calls a Lovable Stack Overflow (LSO), an agent-centric knowledge base of problems and solutions. It holds common configuration issues, database authentication patterns, circular dependencies to avoid, and more. Each entry has two parts: a description that determines when the entry gets retrieved, and a knowledge body that gets injected into the agent's context.

When a user sends a request, a lightweight model checks if the knowledge base already has a solution to that specific problem. If it does, the solution gets injected into the agent's context before it starts working. This system led to a 5% reduction in stuck rates and a 2% higher publish rate in early testing, all without adding latency to the platform.

What Challenges Emerge From Building an AI Knowledge Base?

Building and maintaining LSO has revealed several unexpected challenges. Using AI to frequently update knowledge files tends to degrade into what the team calls "AI slop," so they now focus on finding clusters of similar struggles instead. Knowledge also doesn't age well; an entry that was correct last quarter quietly becomes wrong as software packages evolve. Without aggressive pruning, the knowledge base gets polluted with stale answers and the agent starts injecting outdated solutions.

The team now runs continuous A/B tests by randomly dropping entries from the knowledge base to measure their effect on project success and stuck rates. Entries showing little or negative effect get removed. Additionally, agent failure modes are tied to specific AI models and their training data cutoff dates, so the team must relearn or find new struggles whenever a new model is released.

"We sometimes joke that if the agent were a normal employee it would have complained about its working conditions: unclear instructions and heavy workloads. This sparked the idea: what can we learn by listening to our own agent?" said Benjamin Verbeek, an engineer and former particle physicist on the Lovable agent team.

Benjamin Verbeek, Engineer, Lovable

How Is Lovable Expanding Its Security and Integration Capabilities?

Beyond improving the agent itself, Lovable has been expanding its platform to meet enterprise needs. The company announced a native integration with Wiz, a cloud security platform, that brings security scanning directly into the Lovable build experience. When connected, Wiz scans run automatically as part of Lovable's security suite, and findings surface in the Security view alongside Lovable's existing built-in scanners for dependencies, secrets, database security, and code vulnerabilities.

This integration means that organizations using Wiz can apply their existing security policies to everything built in Lovable without needing new dashboards or extra pipeline steps. For developers, it means dependency vulnerabilities, sensitive data exposure, and configuration issues get flagged in real time. For security teams, applications built in Lovable now appear in their Wiz monitoring dashboard alongside everything else they oversee, providing consistent context and remediation guidance.

Lovable's broader positioning reflects a shift in how non-technical users are building software. The platform has become part of a larger ecosystem where rapid development and governance coexist. By giving its agent a voice through the vent tool and integrating with enterprise security platforms, Lovable is addressing both the friction points that cause users to abandon projects and the compliance concerns that prevent organizations from adopting AI-assisted development at scale.

The company's approach of listening to its own agent's frustrations represents a broader trend in AI development: treating the agent not as a black box but as a source of continuous feedback. As Benjamin Verbeek noted, the goal is to move toward a future where "any issue is experienced only once," turning the agent's complaints into systematic improvements that benefit all users.