Azdaja: a harness-agnostic RLM layer that plugs into Claude Code, Codex, and OpenCode

MiddleDesigner2854 · reddit · 2026-09-19

The author liked Prime Agent's RLM approach but didn't want a full agent harness, so they built Azdaja — a small harness-agnostic layer that plugs into Claude Code, Codex, OpenCode, Pi, etc.

It keeps large inputs in a persistent Python evaluator where code handles search/filter/combine, and only calls the model for parts that genuinely need interpretation. They also experiment with Jev for branches that are really just bounded decisions rather than needing more generated text.

MIT licensed: https://github.com/kubet/azdaja

Original post →

More from coding & agent

coding & agent channel →