The author tested the feasibility of using AI coding agents, such as Codex CLI and Claude Code, for autonomous project development. Initial experiments focused on macOS sandboxing, which failed to prevent unauthorized network access to sensitive data despite blocking direct file reads. The author also encountered issues with child processes continuing to run after the parent process was terminated.
To improve isolation, the author moved to a Linux virtual machine architecture with a broker to manage operations and permissions. While this approach improved control, it did not fully isolate the final product, as JavaScript execution still occurred on the host. Furthermore, the author found that limiting VM memory does not account for total host resource consumption, requiring a separate watchdog mechanism.
The project also addressed system recovery after crashes. The author noted that blindly repeating actions after a failure can cause errors if the previous attempt partially succeeded. A new approach involves tracking the state of external resources and using unique identifiers for connections to prevent conflicts during restarts.
Future development plans include rigorous testing criteria, such as 48-hour and seven-day continuous runs. The author emphasizes that true autonomy requires establishing clear boundaries for data access, process lifecycle management, and reliable state recovery rather than just chaining model calls.
