CODING AGENT HARNESSES

VERIFICATION GATES, SUCH AS TESTS THAT MUST PASS A gate is a checkpoint that work must clear before moving forward. A verification gate is an automatic check whose failure sends the work back instead of letting it through. Think of the security gate at an airport, where it does not matter how confident a traveler feels, the scanner decides. Common gates for coding agents include the following. Unit tests: small tests of individual pieces of code. Linting: an automatic tool that scans code for style problems and likely mistakes. Type checking: a tool that confirms values are used in the ways the code says they should be. Build: confirming the project still compiles or assembles without errors. Security scanning: tools that search for known risky patterns and for exposed secrets, linking to Part Nine. Each gate should have a clear pass criterion and a clear failure behavior. For instance, if the existing test suite does not pass cleanly after a change, the change is rejected and returned to the agent with the failure messages, no matter how confident the agent's own summary sounds. Remember that the agent's statement that it is finished is a claim, not evidence, a point first made in Part Eight. Also remember that gates have limits. Tests only check what they were written to check, so passing tests do not prove the code is good. For high-stakes changes, keep a human reviewer in the loop as well, and track results over time with the evaluation ideas of Part Eight.