Managing a team with AI agents — the playbook for when review becomes the bottleneck

The widespread adoption of AI agents in software development has led to a significant increase in the volume of merged pull requests, yet organizational-level delivery metrics have remained stagnant. This phenomenon indicates that the bottleneck in the development process has shifted from coding to human-centric activities like review, testing, and accountability. AI agents generate code at a pace that far exceeds a human's ability to review, understand, and take responsibility for it, making review throughput the real determinant of a team's pace. To address this, the article suggests a shift in engineering management. Instead of focusing on maximizing agent capacity, teams should cap agent-generated work to what human reviewers can realistically handle. This involves inverting capacity planning to start from review capacity rather than agent count. The analogy of a workshop where many designers produce designs but only one inspector stamps them illustrates this point: the inspector's throughput dictates the workshop's output. Furthermore, the article emphasizes that deploying AI agents necessitates a reevaluation of traditional metrics and practices. Measuring success solely by pull request count is misleading; instead, teams should focus on metrics like churn, defect density, and cost per merge. It also highlights the importance of assigning clear ownership and responsibility to AI agents, treating them akin to employees who require regular performance reviews and management. This ensures that agent outputs are thoroughly verified and that any regressions or drifts in quality are promptly identified and addressed. The core principle is that judgment and verification remain human tasks that cannot be parallelized, making them the ultimate constraint in an AI-augmented development workflow.