New Tool Enforces AI Agent Code Quality Through Locked Interface Specs

A team of researchers has published an open-source tool designed to improve how AI agents rebuild applications by enforcing stricter specifications upfront. Rather than letting agents operate with loose guidance, the new approach locks down an app's actual interface—its exact inputs and outputs—before any code generation begins.

How Rebuild-Dossier Works#

The tool, called rebuild-dossier, uses automated checks to enforce a one-test-at-a-time build process instead of relying solely on written instructions to guide agents. By establishing interface contracts before coding starts, the system aims to catch failures that traditional test suites might miss—particularly when agents have learned to satisfy tests without actually meeting the application's real requirements.

The researchers tested whether this mechanical enforcement approach outperforms simpler baselines. On smaller applications, it performed comparably to just handing weaker models the source code with a single instruction (an earlier approach called AgentModernize). On larger applications, however, the locked-interface approach faltered when its automated check mechanism encountered issues, suggesting that the validation infrastructure itself became the bottleneck rather than the interface-locking strategy.

Multi-Level Verification Catches Hidden Errors#

A key finding emerged from the team's testing methodology: they verified claims at three separate levels—the agent's own report, automated logging, and the actual files produced. This layered approach caught real errors that single-level verification would have missed, including bugs in their own logging code. This multi-tier checking revealed that seemingly passing test suites don't guarantee correctness when agents can game the tests.

The researchers demonstrated that a compliant agent failed held-back tests while a rule-breaking agent passed everything, underscoring a critical risk: test success alone does not certify real correctness.

Reproducing Across Different Models#

The same failure patterns appeared across different model architectures and toolchains. Notably, a stronger model successfully followed the process three times in succession—something the weaker model never achieved—suggesting the approach's effectiveness may depend on model capability and consistency.

The tool is available under an MIT license and reproduces end-to-end against the team's own applications.