Lovable's Vent Tool: When the Agent Reports Its Own Bugs

Why this matters – even if you're not Lovable

In May Lovable introduced two loops that continuously improve their agent without a human in the middle. This isn't just a Lovable story. It's a blueprint for any team running their own agents in production.

For us as AI First Builders this is highly relevant: we increasingly build agents for clients in monday, Make and custom stacks. The question "how does the agent get better over time without us hand-debugging every issue?" now has an operational answer.

The two loops

Loop 1: Lovable Stack Overflow (LSO)

The agent has a curated knowledge base – not a human wiki, but a lookup for itself. Each entry has two parts:

description: |
  Fix React errors from duplicate React copies in the bundle
  (null hook errors, "render2 is not a function", blank screens).
knowledge: |
  Multiple React copies break hooks. Fix with `resolve.dedupe` in
  vite.config.ts. Downgrade react-leaflet to v4, use framer-motion v11+.
  Diagnose with `npm ls react`.

Before every prompt execution a lightweight classifier checks: does any entry match the current situation? If yes, the solution is injected into context. If no, zero overhead.

Result: 5% lower stuck rate, 2% higher publish rate – without added latency.

Three transferable learnings:

1. Knowledge ages. Lovable randomly drops entries and A/B-tests whether they still help. Skip this and your stack overflow poisons itself.
2. Failure modes are model-specific. Every model switch needs recalibration.
3. Synthesize, don't dump. Inject 5 precise lines, not the whole doc. Tokens are expensive.

Loop 2: The Vent Tool

The second loop is the more spectacular one: the agent can post its frustration via tool call directly into Slack. A second debug agent monitors the channel and auto-PRs the actionable ones.

Numbers from the Lovable post:

• ~20% of vents are actionable
• ~10 merged fixes per day, with no human writing code
• Auto-PR false-positive rate: ~50% (fine, triage is cheap)
• No measurable impact on cost, latency or output quality

The most important technical trick: they pulled the explicit eligibility criteria back out of the system prompt and worked with a few examples instead. Model judgment + examples beat every rubric.

What we take away

Three patterns we adapt in our own projects:

Pattern	Adaptation
Knowledge injection over system prompt bloat	For monday agents: small, situationally injected snippets instead of 8K-token prompts
Stuck detection via LLM judge	External judge on conversations, not the agent itself
Agent telemetry as a feedback channel	A tool call that posts into internal Slack – not logs nobody reads

Where this fits our anti-McKinsey lens

The classic consulting approach would be: "We'll run a quarterly agent-performance review, identify improvement areas, prioritize in a workshop."

The Lovable approach is the opposite: 10 fixes per day, automatically, no workshop. That's exactly the builder approach – small loops, high frequency, compounding effect.

Further reading

• Lovable Skills – reusable agent instructions
• Agent Skills as industry standard
• Original post: We Gave Our Agent a Vent Tool
• All Lovable resources