Airlock

Airlock serves as a self-improving controller for delegating and evaluating subagents within coding harnesses. In our own experiences, we've found that the orchestrator agent has the tendency to delegate vague tasks to subagents, leading to overly broad specifications to specific tasks. Furthermore, these subagents evaluate on themselves, writing tests that can evaluate functionality of a feature, yet overlook requirements such as data integrity, ownership boundaries, and safe retry behavior. Airlock delegates approved coding tasks in a schematic, objective-oriented format, which freezes the resulting code, and validates it against independent, controller-owned checks before integration. Changes that fail these tests are rejected, and the subagent must reattempt the fix. Our distinctive feature is improving the validation harness itself, which can be shared across engineering teams. If someone flags an issue with code that passes the harness but fails external tests, the harness will learn to improve its testing ability, until it is able to flag the same issue without overfitting, only accepting a candidate harness if it can detect more defects without reducing the acceptance of legitimate work, thus evolving alongside codebase.