Jazzberry: AI agent for bug finding that thrives in the vibe-coding era by cloning PRs into sandboxes
Jun 11, 2025 with Matteo
Key Points
- Jazzberry's AI agent detects bugs by cloning pull requests into sandboxes and running iterative experiments, targeting the distinct failure patterns LLMs produce when generating code in one shot.
- The startup prices at $20 per developer per month with no founder-led sales, reaching 18 customers with week-over-week growth doubling as it raises at YC Demo Day.
- Jazzberry runs primarily on Gemini Flash and fine-tuned smaller models for tool use rather than frontier models, betting that fast task-specific agents outperform large general-purpose models for sequential code testing.
Summary
Jazzberry builds an AI agent that finds bugs by cloning pull requests into sandboxes and running iterative experiments. When a developer opens a PR, the agent spins up an isolated environment, compiles the code, runs commands, and probes for failures. Each test result informs the next run, making subsequent probes more targeted. The workflow treats QA as a feedback loop: developers generate code in Cursor, Jazzberry tests it for breaks, and results flow back into Cursor.
Vibe coding produces a distinct bug profile. LLMs tend to omit entire function implementations rather than produce line-level typos or variable mismatches typical of human code. Jazzberry's agent targets that pattern, catching missing implementations and changes that were updated in one file but not propagated across the repo.
The agent adapts its testing strategy to the PR. A path-change commit triggers path-focused tests; a scaling change triggers load-oriented testing. Coverage and integration testing are both in scope. Penetration testing is on the roadmap.
Model architecture
Jazzberry runs primarily on Gemini Flash and has fine-tuned a smaller model with reinforcement learning for tool use. Fast, small, task-specific models outperform large general-purpose ones for running many lightweight code experiments in sequence. The founders plan to train further once they map pain points in the current architecture.
Go-to-market
Pricing is $20 per developer per month, with a 7-day free trial accessible through a bot installed directly from jazzberry.ai. There is no founder-led sales motion yet; the landing page serves as the entire go-to-market channel. At the time of the interview, 18 companies were using the product, with the team reporting week-over-week growth doubling over the prior few weeks. One founder worked in LLM-based software testing; the other completed a PhD in reinforcement learning and formal methods. Both are raising at YC Demo Day.