Gubia
- End-to-end automation(
init,phase0,run): planning, implementation and documentation, one checked step at a time. - Built-in self-review: an LLM judge verifies the evidence at each checkpoint and unblocks the loop after repeated rejections.
- Effort matching: every subtask runs at the effort level it needs, with its own list of models per level.
- Fault tolerance: models rotate through a fallback list on failure, and the loop stops with a written report instead of repeating the same failure.
Used in production in a professional data-engineering setting.
Gubia is a meta-harness, a harness-agnostic framework that drives CLI coding agents such as Claude Code, Codex or Devin, and it can be extended with others. Given a goal, it carves it into a plan of small tasks and drains that plan one subtask at a time. It is based on the Ralph Loop, and the methodology page explains why I built on it.
The process is simple. Using Gubia’s skills, you give it a problem and all the context available, and it drafts a plan from them. A phase 0 pass then breaks that draft down into tasks and subtasks, and the loop is launched. From there it is fully autonomous and non-interactive: on each iteration the agent takes the first incomplete task, then the first incomplete subtask inside it, and implements it. Along the way the plan includes review checkpoints, where an LLM judge verifies the evidence, and documentation checkpoints, where what was learned is recorded.
Because the plan is so granular, each task can run at a different effort level. Cheaper models handle the simpler work without losing quality, and the larger, more expensive ones are kept for the tasks that need more reasoning.
If a model fails, the loop moves on to the next one in that effort level’s fallback list, and it stops with a written report if the whole list fails without a single success.
See the repository for the full spec. I built this site with Gubia.
Started May 2026 as a private prototype; public repository since September 2026.