How Specialized Micro‑Agents Cut False Positives by 51% in AI Code Review

Only one key point: instead of feeding a massive context to an LLM, construct multiple small, task‑specific agents.

Cubic (https://www.cubic.dev/), created by former Instagram and Meta engineers, is an AI‑powered code‑review platform whose core feature is an AI review agent that automatically catches bugs, anti‑patterns, and duplicate code in pull requests.

Early user feedback highlighted a major problem: excessive false positives . Even tiny PRs were flooded with low‑value comments, nit‑picking, and obvious false alarms, drowning out useful feedback.

After three major architecture rewrites and extensive offline testing, the team reduced false positives by 51% without sacrificing recall.

1. The “Face‑palm” Phase: A One‑Size‑All Agent

The initial design was straightforward but quickly showed flaws:

Too many false positives – the agent mis‑identified style issues as serious bugs and repeatedly flagged already‑resolved problems.

Loss of user trust – developers began ignoring the comments when half of them were irrelevant.

Opaque reasoning – it was hard to understand why the agent made certain judgments, even when prompts explicitly asked it to ignore minor style concerns.

Conventional tweaks such as longer prompts, adjusting temperature, or sampling strategies yielded little improvement.

2. The Effective Solution

Through extensive trial‑and‑error, the team built an architecture that dramatically cut false positives.

Explicit Reasoning Logs

Require the AI to write out its reasoning before giving any feedback.

{
  "reasoning": "`cfg` may be nil at line 42; line 47 dereferences without check",
  "finding": "Potential nil‑pointer dereference",
  "confidence": 0.81
}

Benefits include clear traceability of decisions, structured thinking that reduces random judgments, and a foundation for diagnosing other issues.

Fewer, Smarter Tools

The original toolchain (LSP, static analysis, test runners, etc.) was trimmed to a simplified LSP and basic terminal, allowing the agent to focus on core problems and improve precision.

Specialized Micro‑Agents Replacing General Rules

Instead of adding countless rules to a giant prompt, the team introduced narrow‑scope agents:

Planner : quickly assesses changes and selects needed checks.

Security Agent : detects injection and authentication vulnerabilities.

Duplication Agent : flags duplicated or plagiarized code.

Editorial Agent : handles spelling and documentation consistency.

This specialization keeps each agent focused, uses tokens efficiently, and boosts accuracy, with the trade‑off of some context overlap managed via caching.

3. Real‑World Results

Across hundreds of open‑source and private repositories over six weeks:

False positives dropped by 51%, restoring developer trust.

Average comments per PR halved, letting teams concentrate on truly important issues.

Review flow became smoother, with faster merges and less time spent on irrelevant feedback.

The noise reduction markedly increased developer confidence and participation.

4. Key Takeaways

Explicit reasoning improves clarity : forcing the AI to explain its rationale raises accuracy and simplifies debugging.

Streamline the toolchain : regularly prune tools used by less than 10% of cases.

Micro‑agent specialization : assign each AI agent a single, narrow task to reduce cognitive load and boost precision.

These lessons apply not only to code‑review agents but to the design of any AI agent.

Original source

Learnings from building AI agents (https://www.cubic.dev/blog/learnings-from-building-ai-agents)