Bonzai Dashboard
Blog · Case study

How Bonzai Uses Bonzai to Find and Fix Its Own Production Bugs

In its first week watching its own product, Bonzai found 76 problems in Bonzai and opened 46 fixes. The team merged all 46, 40 of them exactly as Bonzai wrote them.

Summary

  • Seamless Integration: Bonzai integrated into its existing infrastructure without adding tools, using the team’s existing feedback channels: Logfire, Sentry, Vercel, and ClickHouse.
  • Merged by a Person, Tracked After Release: A person merged each of the 46 PRs. Bonzai’s PRs accounted for 35% of all PRs the team merged into production. After release, two fixes were confirmed in production, and one (#326) regressed.

About Bonzai

Bonzai monitors a product’s production signals, detects abnormal behavior and errors, and opens pull requests with evidence of the bug and its fix. While most products take a reactive approach to finding and fixing problems, Bonzai does both: it proactively scans the codebase and creates fixes before telemetry flags them, and it finds problems in the telemetry itself.

Bonzai also integrates into a team’s codebase without requiring a migration to another logging platform. Teams often don’t want to pay the opportunity cost of adopting another platform, and doing so is often futile. Bonzai connects to the tools teams already use for errors, logs and traces, product analytics, hosting, and databases. Its catalog includes Sentry, Datadog, Grafana, Honeycomb, New Relic, Logfire, PostHog, Vercel, Cloudflare, Google Cloud Run, ClickHouse, and App Store Connect. Its investigations can run on any Anthropic or OpenAI models; this account used Sonnet 4.5.

Bonzai’s cloud backend, website, and a cloud agent for every customer account all ship from one repository. On November 11, 2026, the team connected that repository and its production tools to Bonzai, letting Bonzai monitor itself.

Our goal for using Bonzai:

  • Find problems before anyone reports them - Go beyond reacting to bug reports and alerts: look for what is failing in production, including failures that never raise an error.
  • See how it reached each fix - Keep every query it ran, every claim about a cause, and the source behind each claim on record and open to review.

How the Bonzai team used Bonzai

Bonzai’s connection to Logfire, the team’s log and trace store, was the main source. Bonzai queried it directly with SQL: 1,843 of its 2,490 queries were SQL against Logfire. It also read Sentry issues, Vercel deployments and ClickHouse. Of the 76 problems it found,

  • 51 came from Logfire
  • 2 from Sentry
  • 10 from a scan of the code for errors that are caught and dropped
  • 9 from reports by Bonzai’s own agents
  • 3 from dependency and security alerts
  • 1 from a flaky deploy check

Bonzai classified 46% of the problems as errors and 32% as silent failures: issues where something went wrong without being reported. The rest involved flaky tests, performance, prompt issues, users retrying or giving up, security, and dependencies.

Every problem received a priority score, and 49% also had a written reproduction. Before triage, Bonzai investigated each problem, grounding its findings in feedback from analytics tools, the recorded intent, or the codebase.

In conversations with customers, we’ve heard that AI-assisted coding tools often treat trivial errors as equally important and create unnecessary fixes. Bonzai addressed this by dismissing 13 of the 76 problems as requiring no fix, with a concrete justification for each dismissal.

After finalizing the issues, Bonzai put each fix behind a feature flag and rolled it out to production without human intervention. A retrospective review of the PRs found no major concerns, so no rollback was needed

Over the seven days from November 11 to 17:

Problems found 76
Fix pull requests opened 46
Merged 46 of 46, each by a person
Merged with no changes at least 40 of 46 (87%)
Share of all PRs merged into the repository 46 of 132 (35%)
Problem found → PR opened, median 6.3 hours (34 PRs linked to a problem)
PR opened → merged, median 5.2 hours
  • Failures nobody could see (7 PRs): Errors that never reached Sentry or Logfire, so nobody was told.
  • Failures with no explanation (12): Failures that were logged without the reason, step, release or output needed to act on them.
  • Connections to other tools (10): Reading PostHog, Sentry, ClickHouse, App Store Connect and Cloud Run broke, or failed without saying why.
  • Bonzai’s own engine (8): Bugs in its code map, history import, duplicate checks, task queue and log intake.
  • Cloud agent updates and restarts (6): Updates, relaunches and task starts that failed or blocked.
  • Test tooling (2) and speed (1): One fix removed quadratic time from fetching pages of results.

What’s next

  • Video proof of a fix, next to the cited evidence Bonzai writes today.
  • Conversion and churn. Finding where people drop off or leave, from the team’s product analytics.

Does your team ship faster than it can watch production?

Bonzai connects to the logging, error and analytics tools you already have and opens fixes with the evidence attached, for your team to review and merge. Get started at hibonzai.com.