Lovable’s AI agent can now complain about its job. The team built a send_feedback tool — internally called the “vent tool” — that lets the agent flag problems it encounters mid-task: unclear instructions, impossible requests, missing context. Those complaints feed directly back to the engineering team and into a growing corpus of problems and solutions.
The practical mechanics:
- The agent calls
send_feedbackduring production runs when it hits friction — ambiguous prompts, conflicting requirements, tasks it can’t complete cleanly - Feedback builds a structured corpus of real failure modes, paired with how the agent attempted to resolve them
- The system self-improves hourly, using that corpus to patch its own behavior against recurring issues
- Engineers review the vent logs to find patterns — the agent effectively triages its own bug reports
Lovable cofounder Fabian Hedin presented the concept at Code with Claude, framing it as giving the agent “a venting mechanism for when it’s having a bad day.” The joke lands because the underlying problem is real: when 100,000 projects are being built per day on your platform — a number Lovable hit alongside $100M ARR in 8 months — the agent encounters every possible flavor of bad prompt. Manual QA can’t keep up.
The team compared the agent to an employee who’d complain about working conditions if it could. So they made it possible. The difference from standard error logging: send_feedback captures subjective friction, not just stack traces. The agent can flag that a prompt was technically parseable but practically nonsensical — the kind of signal that traditional monitoring misses entirely.
NOTE
The vent tool doesn’t pause or refuse work. The agent completes the task as best it can, then files its feedback asynchronously. Users don’t see the complaints.
Lovable is processing 100,000 projects daily. The vent tool turns every frustrating prompt into training signal — a feedback loop where the agent’s complaints make tomorrow’s agent marginally less frustrated. The blog post frames this as “self-improving every hour,” which is a strong claim, but the architecture is sound: production friction as structured data, fed back continuously.