TriageAI: From Bug Report to Root Cause in 60 Seconds
Every engineering team knows the routine. A bug report lands in Jira or Sentry, someone reads the stack trace, opens the file, scrolls to the failing line, runs git blame, scans the recent commits and finally works out what changed. Only then does the fix start.
TriageAI automates that routine. It's a Claude AI agent for debugging that takes any bug report and returns three things in under a minute: the root cause with the exact file and line, a suggested fix as a diff, and a regression test.
The problem: triage is slow, and it's the same work every time
Finding the root cause is usually the longest step in fixing a bug. We estimate it takes 45 to 90 minutes per bug, and it lands on whoever knows the codebase best. That has three costs:
- Time. An hour of investigation before a one-line fix.
- Blocked juniors. Engineers new to a codebase wait for a senior engineer, who then gets pulled off their own work.
- Missing tests. When the fix is rushed, the regression test is the first thing to go, and the bug comes back.
What TriageAI does
TriageAI is AI bug triage that ends in something you can act on. For every report it produces:
- Root cause. One paragraph naming what broke, why, and the file and line responsible. If a recent commit introduced the regression, it names the commit.
- Suggested fix. The minimal change, as a before/after diff, with a note on why it's safe.
- Regression test. A complete test in your codebase's language and framework that fails before the fix and passes after it.
It accepts Jira tickets, Sentry errors, GitHub issues and raw log pastes. With the GitHub webhook connected, every newly opened issue is triaged automatically and the diagnosis is posted back as a comment.
How it works
TriageAI has five layers.
Input sources and normalization. Each source has its own format, so the first step converts every report into one shape: error type, message, stack trace and context. A sanitizer strips secrets and personal data at the same time.
The agent loop. TriageAI calls Claude with a system prompt, the normalized report and a set of tools. Claude decides what to investigate. While the response stops with tool_use, TriageAI runs the tool, appends the result and calls Claude again. When it stops with end_turn, the investigation is done. A cap of 12 iterations and a timeout keep it bounded.
call Claude(system prompt, report, tools)
while stop_reason == "tool_use":
run the requested tool
append the tool_result
call Claude again
format the answer and post it to the ticket
The tools. Claude gets five, all read-only and scoped to the repository:
search_codebasefinds files, functions or patternsread_filereads a file or a range of linesgit_blameshows who last changed a line rangelist_recent_commitsshows what changed recently in a file or directoryrun_testruns an allowlisted test command in a sandbox
The system prompt tells Claude to start from the top of the stack trace, read the code before drawing conclusions, check blame and history, and never make a claim it can't trace to a file, line or commit it actually read. If it can't reach a confident answer, it says so.
Formatting and posting. The final answer is split into its three sections and posted to the GitHub issue, the Jira ticket or the web UI, along with the full investigation trace.
Impact
These are estimates, not measured customer results:
| Metric | Before | With TriageAI |
|---|---|---|
| Time to root cause | 45–90 min | ~60 sec |
| Fixes with a regression test | ~30% | 100% |
| Junior self-sufficiency | Low | High |
| Senior interruptions | Frequent | Minimal |
For a 10-person team, that adds up to an estimated 20+ hours reclaimed per sprint. It's also a direct way to reduce MTTR, because the slowest step in the incident is the one that gets shorter.
Security best practices
An agent that reads your code has to be careful with it. TriageAI is built around these practices:
- Secret redaction. API keys, tokens, passwords, connection strings and emails are stripped before input reaches the model.
- Read-only, scoped access. File tools work on a read-only clone and reject any path outside it.
- Audit logging. Every tool call is logged with its input, output, timestamp and request ID.
- Human review. Output is labeled as a hypothesis to review. Nothing is ever merged automatically.
- Sandboxed execution.
run_testonly runs allowlisted commands, with a timeout and no network. Report text can never choose a command. - Prompt-injection awareness. The bug report is treated as untrusted data, and instructions inside it are ignored.
- Key hygiene, verified webhooks and limits. Keys stay in server environment variables, webhook signatures are verified, and the triage endpoint is rate limited and size capped.
Read the full list on the security page.
What's next
On the roadmap: multi-language support for Java, Python, Go and TypeScript, two-way Jira sync, Slack alerts on new Sentry events, and a team knowledge base of past diagnoses.
If your team spends its mornings reading stack traces, see how TriageAI works.