Skip to content

yapusStandards and orchestration for coding with an AI agent

If an AI writes the code, what keeps it sound?

Features ​

  • Quality held by checks. Standards are written files; a rule a program can check is held by a hook or a test, and a pass-or-fail command, not a sentence, decides that a change is done.
  • Independent review. A plan and a finished diff are judged by a subagent that never saw the conversation that produced them.
  • Context management. The main session keeps what a decision needs; subagents do the reading, running and editing, each return is capped, and a standard other than the principles loads only when a file of its kind is touched.
  • Parallel work. Independent tasks launch together, each subagent on its own files.
  • Work sized to the task. A small change runs with no plan; a large one gets a written plan, built and committed one part at a time.
  • Survives compaction. The prompts the person typed are logged and given back to the session after a compaction, with the paths of the newest plans.
  • Cost control. Each subagent runs on the model its definition names, a smaller one for running commands, reading and web research than for editing, and a session has a launch budget.