Steerability via constraints: a substrate for scalable oversight of coding agents

As AI coding agents become more capable, the primary challenge for developers is maintaining effective human oversight. Unconstrained agents can introduce security risks and make codebases difficult to manage, leading to expensive and time-consuming review processes. This paper proposes a solution by applying traditional software engineering management techniques—such as access control, network policies, and strict coding conventions—directly to AI agents. By creating a "constrained substrate," the author argues that we can improve oversight efficiency and security more cost-effectively than through current agentic scaffolding methods.

Applying engineering discipline to AI

The core idea is that the methods used for decades to manage large human engineering teams are highly effective for AI agents as well. Instead of relying solely on complex agentic frameworks, the author suggests that enforcing strict rules at the substrate level—the environment where the agent operates—provides a more reliable foundation. By implementing tools that enforce coding standards and restrict network or system access, the environment itself acts as a guardrail, ensuring that the agent’s output remains within safe and predictable boundaries.

Improving security through constraints

To test this approach, the author conducted an experiment using a small AI reviewer (Gemma 4 e4b) to inspect a Python codebase containing 11 hidden backdoors. The results demonstrated a significant improvement in detection capabilities when constraints were applied. Without any tools or constraints, the agent’s recall rate was 54.5%. When the agent was provided with a constrained substrate combined with a simple 200-line documentation CLI tool, the recall rate increased to 90.9%. The study notes that both the substrate constraints and the additional tools contributed to this success independently.

Why Python and future scalability

The author chose Python for this experiment because it is a language that offers fewer built-in guarantees compared to others, making it a prime candidate for demonstrating the benefits of substrate-level oversight. By imposing structure on a flexible language, the system can catch issues that might otherwise go unnoticed. While the experiment focused on Python, the author suggests that these principles are universal and can be extended to more strictly typed languages like Rust, providing a scalable path forward for managing AI-driven software development.