What Is a Quality Gate? Definition, Criteria, and Common Mistakes

A quality gate is a defined checkpoint with fixed criteria that a change must pass before it is handed on. It is not a recommendation and not a metric, but a barrier: criteria met means proceed, criteria not met means stop. This clarity is the entire value: a gate that can be bypassed is just a report.
What Is a Quality Gate?
The term comes from project management and originally referred to a milestone at which defined deliverables had to be in place before the next phase could begin. In software development, it has taken on a narrower, technical meaning: an automated check in the delivery chain that aborts the process if it is not met.
Three properties make a gate a gate:
- The criteria are fixed in advance. Not negotiated case by case, but formulated as a rule.
- The decision is binary. Passed or not. "Borderline, but it's urgent" is not a valid answer for a machine to give.
- It is binding. A failed gate stops the process. Otherwise it is just a notice.
Which Criteria Belong in a Quality Gate?
Only criteria that a machine can decide unambiguously. Everything else belongs in review, not in the gate.
| Criterion | Sensible threshold |
|---|---|
| Test run | no failing tests, no exceptions |
| Coverage of changed lines | not below the existing project value |
| New static analysis findings | none above the defined severity |
| Known vulnerabilities in dependencies | no new critical ones |
| Secrets in the source code | none, a violation always aborts |
| New duplicates | none above the defined block size |
The decisive distinction is in the second column: the thresholds apply to new and changed code, not to the existing codebase. A gate that evaluates the legacy codebase blocks everything on day one and gets switched off on day two.
Where in the Workflow Should a Quality Gate Sit?
At several points, with increasing strictness and decreasing frequency.
- At commit time (seconds): formatting, obvious errors, secrets. Whatever fails here costs nothing.
- At the pull request (minutes): tests, static analysis, coverage, dependencies. The actual gate.
- Before deployment (the hard barrier): everything before that, plus the questions that only matter here: is there a migration involved, is it reversible, is there a rollback path, was something included that should never end up on a server?
The last point is often forgotten because it does not look like code quality. It is the most expensive one: a wrongly shipped configuration file or an irreversible migration is the only mistake on this list that can take a system down for hours. Details in Quality Gates in the CI/CD Pipeline.
Why Do Quality Gates Fail in Practice?
Because of four mistakes that all have the same effect: the gate gets bypassed instead of used.
- Too many criteria at once. Anyone who starts with fifteen rules blocks every single process and gets switched off within a week. Start with three.
- Evaluating the existing codebase instead of what's new. The most common setup mistake, see above.
- Bypassing without a trace. There has to be a way to override a gate, otherwise the first special case blocks operations entirely. This override must be logged and carry a name. A silent bypass makes the gate worthless.
- Criteria that never decide anything. A gate that has been green for a year straight is not checking anything. Either the threshold is too lax or the criterion is irrelevant to the current way of working.
Why Have Quality Gates Become More Important for AI Code?
Because review capacity is not keeping pace with generation speed. A person does not review more lines per hour today than three years ago, but significantly more are being generated.
The result is a shift that a person cannot see in a single diff: more duplication, less cleanup (figures and where they come from), and a threshold in the pipeline reports it immediately. This is exactly what a gate is the right tool for: it does not compare the line, it compares the state.
Frequently Asked Questions About Quality Gates
What Is the Difference Between a Quality Gate and a Definition of Done?
The gate checks mechanically and decides in binary terms. The Definition of Done is the team's agreement and also includes points that no machine can judge. The gate enforces the technical part of the Definition of Done; see Definition of Done.
Do I Need SonarQube for a Quality Gate?
No. Any CI pipeline can abort on a failure, and that is already a gate. A platform provides history, threshold management, and reporting, but it is not a prerequisite.
How Many Criteria Should a Gate Have?
Three to start: no failing tests, no new critical analysis findings, no secrets in the code. Expand once these three are consistently green.
What If a Gate Blocks Urgent Work?
Then it is overridden with documentation, a name, and a reason, and the case is followed up afterward. If bypasses pile up at the same spot, the criterion is wrong, not the urgency.
A gate only works if it stands in front of the step, not next to it. That is exactly how Code Guardian is built: seven gates for deployment, migration, new dependencies, data conclusions, findings, decision questions, and legal text, six of them backed by a hook that actually halts the process.
Sources
- GitClear, The Maintainability Gap: 2026 AI Code Quality Research.
- ISO/IEC 25010:2023 (quality characteristics that gate criteria can be aligned with).
As of: July 27, 2026. General professional overview.
Share this article
Stay up to date
Get the latest articles, insights and industry updates straight to your inbox.
Decide for yourself what Google shows you
Google lets you choose which sources appear more prominently in your search results: in Top Stories and in AI answers. Two clicks, and you see the sites you trust.
Add provimedia.de to my preferred sourcesRelated articles
More articles you might find interesting.
Who Is Liable for AI-Generated Code? What Changes in 2026
As of December 2026, software is explicitly a product under product liability law. What this means for developers, agencies, and clients.
Definition of Done for AI Code: Evidence Instead of Promises
When code is written in minutes, the bottleneck shifts to the question of when something is actually done. A Definition of Done built for that pace.
Code Audit: A Checklist for Reviewing Someone Else's Software
Before a handover, before a purchase, after an incident: what a code audit checks, in what order, and how to spot a superficial one.
Prüfungen, die sich nicht überspringen lassen
Das Skill-Paket für Claude Code und OpenAI Codex: sieben Gates vor Deploy, Migration, neuer Abhängigkeit, Datenurteil, Befund, Optionsfrage und Rechtstext. Firmenlizenz, unbegrenzt viele Entwickler im Unternehmen.