A study demonstrates that applying traditional engineering controls, such as access policies and strict conventions, to coding agents significantly enhances scalable human oversight compared to unconstrained approaches.

  • Researchers implemented a constrained substrate with a ~200-line `docs` CLI for Python codebases containing 11 inserted backdoors.
  • A small reviewer model (Gemma 4 e4b) achieved a recall rate of 90.9% using the constrained system, up from 54.5% in unconstrained scenarios without tools.
  • The experiment showed that both the substrate and the tools contributed independently to the improvement in detection capabilities.

This approach offers a cheaper alternative to recent agentic scaffolding methods by leveraging established team management principles for automated oversight.