The video advocates for “meta automation” by embedding deterministic, domain-specific rules into automated tools like custom ESLint rules to enforce coding standards, rather than relying on multiple AI agents reviewing each other, which adds complexity and inefficiency. This approach improves code quality, saves time, and future-proofs projects by enabling AI agents to self-correct within clear, testable frameworks, ultimately promoting better engineering discipline in AI-assisted development.
The video explores the challenges of ensuring good engineering practices when working with AI agents that occasionally produce inconsistent or subpar work, likened humorously to colleagues showing up drunk. The presenter shares insights from research with early adopters of AI agentic development, emphasizing the importance of “meta automation”—the automation of the automation process itself. Instead of relying on multiple layers of agents reviewing each other, which adds complexity and inefficiency, meta automation focuses on creating deterministic, automated feedback loops that guide agents to produce code aligned with team policies and standards.
A common approach today involves defining agent guidelines in markdown files, often called skills, rules, or constitutions. However, these are insufficient because agents sometimes ignore them, leading to the problematic cycle of adding more reviewer agents. The presenter argues this approach is flawed and instead advocates for embedding contextual, domain-specific rules into deterministic tools that operate independently of the agents. For example, custom linting rules in ESLint can enforce critical coding standards, such as restricting network calls to a specific runtime class, which agents can then check automatically before human review.
The video highlights a practical case where the presenter used an AI agent to research, write, and test custom ESLint rules that detect unauthorized network calls. This meta automation caught errors missed by both human and agent reviewers, demonstrating how automating rule enforcement within the build process saves significant time and reduces mistakes. The approach also includes writing clear, actionable error messages that help agents self-correct without human intervention, preventing endless loops of failure and review.
Importantly, the presenter notes that while markdown rules are a good starting point, teams should transition to deterministic, testable tools as soon as possible. This shift not only improves consistency and reliability but also future-proofs the project by decoupling rule enforcement from any specific AI agent or framework. The ability to automate and maintain these custom rules at near-zero cost thanks to AI agents democratizes process improvements that were previously only feasible for large organizations with dedicated developer platform teams.
Finally, the presenter invites viewers to participate in a survey to help validate these findings and improve community guidelines for working with AI agents. By sharing experiences, developers can contribute to refining best practices and receive aggregated results with personalized recommendations. The overall message encourages applying proper engineering discipline through meta automation to save time, improve code quality, and onboard new team members more effectively when working with AI-assisted development.