What Is a Software Regression? Causes, Detection & Prevention

The canonical definition comes from the testing-standards body. ISTQB defines a regression as 'a degradation in the quality of a component or system due to a change.' Read that sentence carefully: the trigger is a change, and the failure is degradation of something that already existed. That is the whole distinction. A net-new bug is behavior that never worked; a regression is behavior that worked yesterday and does not work today, which means there is a prior version to compare against.

That baseline is not a footnote. It is the single most useful property of a regression. It is why git bisect exists. If the feature passed on commit A and fails on commit Z, the breaking change is somewhere between them, and you can binary-search the range to find it. A brand-new bug gives you no such anchor. So the first question to ask of any defect is not 'where did it throw' but 'did this ever work', because the answer routes you to two completely different debugging strategies.

Regressions matter because they are the failure mode that scales with how fast you ship. The production-facing version even has a metric: DORA tracks change fail rate (the share of deployments that cause a failure in production requiring a hotfix, rollback, or patch) as one of its four core delivery measures. That makes the regression a benchmarkable number, not a vibe. If your change fail rate is climbing, your safeguards are losing the race against your deploy frequency.

And the substrate they surface from is enormous. CISQ put the cost of poor software quality in the US at $2.41 trillion, including roughly $1.52 trillion of accumulated technical debt. Technical debt is precisely the brittle, under-tested code from which regressions repeatedly emerge whenever someone touches it. The harder problem is that detection is a needle-in-a-haystack search: in Google's 'Taming Google-Scale Continuous Testing' study, only 1.23% of test executions actually caught a real breakage, while about 84% of observed pass-to-fail transitions were flaky tests rather than genuine regressions. Most of the signal your suite produces is noise, which is why a stable baseline and reliable tests matter more than raw test count.

Here is the honest limit of a stack trace on a regression, and where BugMojo fits. A trace tells you where code failed and the path that got there. But a regression's root cause is the change that altered behavior, and that change is frequently nowhere near the line that threw: a config flag, an upgraded dependency, a different API response shape. The failing code is not necessarily the code that changed. So a clean trace points you at a symptom while the cause sits three files and one deploy away.

In a BugMojo report the trace does not arrive alone. The browser extension captures the failure with its surrounding state (an rrweb session replay, the console output, and the network requests around it, with method, URL, status and timing), so the frame at Pricing.tsx:142 sits next to the GET /api/plan call that came back just before it threw, and the replay shows the plan page the user saw. That is the difference between 'something broke near line 142' and 'it broke right after the plan endpoint answered, so check what that endpoint now returns.' The state is what tells you a change caused it.

Then BugMojo hands that whole bundle to an AI agent (Claude Code, Cursor) over an MCP server. The agent reads the replay, the console, and the network requests together with your repository, so it can correlate the failing state with the diff that introduced it instead of guessing from a trace.