Quality assurance · Production

How cubic reduced AI code review false positives by 51% with specialized micro-agents

The problem

Cubic's AI code reviewer generated excessive low-value comments, nitpicks, and false positives, causing developers to lose trust and ignore the feedback altogether, obscuring genuinely valuable findings.

First attempt

The initial do-everything single-agent architecture with extensive tooling produced excessive false positives and opaque reasoning. Standard remedies—longer prompts, temperature adjustments, and sampling experiments—had minimal effect.

Workflow diagram · grounded in source
1
PR triggers code review
trigger
“an AI code review agent that performs the first review on a PR, catching bugs, anti-patterns, duplicated code, and similar issues”
2
Planner identifies checks
ai_action
“Planner: Quickly assesses changes and identifies necessary checks”
3
Specialized agents analyze
ai_action
“Security Agent: Detects vulnerabilities such as injection or insecure authentication. Duplication Agent: Flags repeated or copied code. Editorial Agent: Handles typos and documentation consistency.”
4
AI states reasoning before feedback
validation
“We required the AI to explicitly state its reasoning before providing any feedback”
Reported outcome

After three major architecture revisions, cubic reduced false positives by 51% without sacrificing recall and cut the median number of comments per PR by half, resulting in smoother review processes and improved developer trust.

Reported metrics
False positive reduction51%
Median comments per pull requestcut by half
Reported stack
Language Server Protocol (LSP)static analysistest runnersterminal
Source
https://www.cubic.dev/blog/learnings-from-building-ai-agents
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

After three major architecture revisions, cubic reduced false positives by 51% without sacrificing recall and cut the median number of comments per PR by half, resulting in smoother review processes and improved devel…

What tools did this team use?

Language Server Protocol (LSP), static analysis, test runners, terminal.

What results were reported?

False positive reduction: 51%; Median comments per pull request: cut by half (source-reported, not independently verified).

What failed first in this deployment?

The initial do-everything single-agent architecture with extensive tooling produced excessive false positives and opaque reasoning.

How is this quality assurance AI workflow structured?

PR triggers code review → Planner identifies checks → Specialized agents analyze → AI states reasoning before feedback.