AI Supervising AI
It is one of the components of the reinforced DoD for AI-enabled teams and part of the team's AI governance system.
Why it exists
When an AI agent generates code, documentation, or analysis, there are quality aspects that can be verified automatically and quickly: syntax errors, known insecure code patterns, duplication, and non-compliance with style conventions. Having a human review these mechanical aspects consumes valuable time that could be devoted to reviewing the business logic.
"AI supervising AI" delegates these mechanical checks to another agent, freeing the product builder to focus on the aspects that only human judgment can verify.
Common use cases
- Review of generated code: an agent analyzes the code produced by another agent, looking for known vulnerabilities, problematic patterns, duplication, and compliance with the project's style.
- Spec verification: an agent verifies that a spec is complete, non-contradictory, and covers the edge cases defined in the project.
- Hallucination detection: an agent with access to the official documentation of the project's APIs verifies that the generated code uses functions and methods that actually exist.
- Test review: an agent verifies that the generated tests cover the scenarios described in the acceptance criteria.
Limits of the mechanism
"AI supervising AI" is effective for checks that can be formalized as rules. It cannot replace human review in:
- Validating business logic in context.
- Detecting incorrect assumptions about the product domain.
- Assessing whether the code is maintainable by the team in the long term.
- Architecture decisions that require understanding the system as a whole.
An agent can detect that a deprecated function is being used, but it cannot detect that the algorithm is solving the wrong problem.
Risk of error cascade
When the reviewing agent and the generating agent share the same underlying model or the same assumptions, they may have the same blind spots: the reviewing agent fails to detect what the generating agent would not detect either. This risk is reduced by using different models or configurations for generation and review, or by using deterministic static analysis tools (not based on language models) for the most critical checks.
Common mistake
Assuming that "AI supervising AI" is equivalent to human review. They are not equivalent: they are complementary. The complete review chain in a healthy AI-assisted development flow includes automated review by static tools, agent review, and human review. Removing human review because "the AI has already reviewed it" is the most common source of accumulated technical debt in teams that adopt AI at an accelerated pace.
Resources
🏦 Scrum en equipos con IA [Scrum in Teams with AI, in Spanish]Skill Arena · Scrum Manager
📊 Guía didáctica Scrum en equipos con IA [Teaching Guide: Scrum in Teams with AI, in Spanish]Resources · Scrum Manager
See also
Want to advance in agility? You can search for course and exam calls or go at your own pace by becoming a member of the Agile Club. This membership includes exclusive resources, e-learning classrooms, and access to Skill Arena: a space to practice and measure your agile skills at your own pace.