Descriptions:
Kyle Mistele, co-founder of HumanLayer, presents a principled engineering framework for building AI coding loops that work in production — not just in demos — at the AI Engineer conference. His central argument is that the industry’s current approach of piping prompts into unbounded loops produces unreadable 40,000-line PRs that fail under real-world constraints like team collaboration, regulatory obligations, and service-level agreements.
Mistele grounds his framework in control theory: the same discipline used in aerospace engineering for flight stabilization. In his model, a sensor measures the current state of the codebase against a desired end state, a controller determines what incremental change to make, and an actuator applies it — with the loop repeating until the system converges. He introduces ASGrep as a language-agnostic sensor tool that identifies code patterns across polyglot monorepos without being disabled by inline TypeScript comments or ESLint overrides, a known failure mode when Claude is given unconstrained access to a codebase.
As a concrete worked example, Mistele walks through HumanLayer’s internal loop for incrementally migrating their RPC API to the Effect library. The sensor identifies unmigrated procedures via ASGrep rules; the controller applies targeted, reviewable changes one procedure at a time rather than in a single sweeping rewrite. The talk argues this loop-plus-control-theory approach is how teams can move fast with agents while still reading the code they ship.
📺 Source: AI Engineer · Published July 25, 2026
🏷️ Format: Deep Dive
