Humans out of the loop (matthewboston.com)

A recent discussion in the AI/ML community highlights the need to reconsider the "human in the loop" approach when handling AI agents, particularly regarding trivial changes that often lead to unnecessary bottlenecks. The author argues that relying on human approval for every minor adjustment — such as dependency bumps or configuration changes — can turn reviewers into "meat proxies," merely clicking through without meaningful oversight. This practice not only adds to the reviewer’s workload with little to no real benefit but can also create false security in the review process. Instead, the article proposes a more streamlined approach that emphasizes establishing clear guardrails, allowing certain changes to bypass human review if they meet predefined criteria. These criteria include good testing practices, small diffs, canary deployments, fast rollback capabilities, and explicit definitions of what constitutes a trivial change. By shifting the review responsibility from human approval to automated systems, the goal is to enhance trust and efficiency in deploying changes while freeing up human reviewers to focus on more critical decisions. In doing so, teams can maintain quality without overwhelming their members with trivial tasks.