Engineering Practice

• We Built Ourselves an AI Coding Workbench — Using GitHub Actions as the Sandbox

  Why we passed on a self-hosted K8s sandbox cluster, third-party sandbox services, and off-the-shelf managed agents — and ended up using GitHub Actions runners as the sandbox for our team's AI coding workbench. A record of the design trade-offs: treating GHA's instability as part of the contract, making conversation continuity the only hard guarantee, and using OIDC to keep zero secrets in the sandbox repo. Many technical decisions only become available once you stop trying to do everything.

  May 8, 2026 - 14 minute read - AI-Assisted Development GitHub Actions Infrastructure Agent Engineering Practice
• "All Tests Pass" Doesn't Mean "Requirements Are Met" — Why AI-Assisted Development Is Often Suspiciously Green

  When AI-assisted long tasks end with all-green CI but still ship bugs, the problem isn't AI laziness — it's that engineering signals were never designed to detect silent passes. This piece dissects the "green but wrong" anti-pattern: five forms, five mechanisms, and nine structural countermeasures.

  Apr 27, 2026 - 11 minute read - AI-Assisted Development Testing Software Engineering Anti-Pattern Engineering Practice