Kerr argues team mental model sync is the AI bottleneck
Dave Kerr contends that the primary constraint in AI-assisted engineering is not model capability but the team's ability to keep its mental models synchronized with system reality. He advocates for deliberate slowdowns through rigorous review processes to maintain architectural integrity.
Dave Kerr, creator of the Hacker Laws repository and author of Effective Shell, argues that the limit to how fast teams can build and scale is how well they keep their mental models in sync with reality. While AI agents can one-shot complex systems, Kerr notes that these outputs often lack reusable primitives and violate design principles such as least astonishment. To counter this, his team deliberately slows down using spec reviews and code reviews, ensuring engineers understand the architecture before accepting agent-generated code.
Kerr also highlights a bias in large language models to build new solutions rather than using existing hardened tools like Bash. His team spends 25% of its engineering time dealing with CVEs in the NPM ecosystem, a burden he attributes to this build-over-buy tendency. To manage agentic workloads, Kerr advocates for using Kubernetes as if the agents were microservices, leveraging its resource management, audit logging, and GitOps capabilities. His team uses an open-source tool called Arc to define agents in YAML, which allows for version control and governance in regulated environments. Kerr maintains that these deliberate slowdowns are necessary to preserve system integrity and prevent the acceptance of complex code without proper scrutiny.