HumanLayer co-founder Kyle Mistele argues that enormous unattended pull requests are a poor fit for many teams. He uses control theory to describe an alternative: measure the current state, compare it with a desired state, choose an incremental action and measure again. Deterministic checksDeterministic software verification checks an output with a repeatable procedure that produces the same result for the same inputs and state. can supply a reliable sensor while an agent handles a narrowly scoped changeAn AI coding agent is a tool-using AI system that can inspect, modify, and validate software within a repository..
A code-migration example combines AST-grep checks with a recorded baseline, hand-written migration patterns and a skill for producing consistent changesAn AI agent skill is a reusable package of instructions, resources, and tool guidance for performing a bounded kind of work.. Existing CI detects regressions, and the controller selects manageable work instead of asking the model to transform the entire codebase at once.
Human review becomes feedbackAn AI feedback loop occurs when AI outputs or their consequences become inputs that influence later model or agent behavior. for subsequent iterations. Limiting each loop to one open pull request prevents a backlog of unreviewed changes; separate bounded contexts can support larger programs of work. Mistele presents this as a practical engineering approach, not a guarantee that any autonomous coding loop will be safe.
Watch on YouTube




