JevHarness lets an LLM author and freeze task-specific agent harnesses — but who owns the frozen SOP?

sujingshen · x · 2026-09-22

The open-source JevHarness project (96 stars on GitHub) lets an LLM write a task-specific decision harness once — auto-generating features, questions, criteria and action logic — then freeze it so execution runs only code plus fast Jev fuzzy judgments, no per-step LLM calls. It supports reward-based full-trajectory reflection / GEPA self-optimization; the demo shows Pokémon battle win rate going from 25% to 75% in 5 iterations, with a Claude Code skill install path and replayable demos.

sujingshen adds a deeper concern: ownership of the frozen strategy. Who wrote the frozen "window card", who can change it during "continuous evolution", and can you keep your veto history when it errs? If everything is locked in someone else's default pipeline, switching models makes the card evaporate — a great win rate with nowhere to place your red lines. He proposes four checks: can the pre-freeze instructions/criteria/candidate sets be exported; is human review with a stop switch in place; is there a veto history; and does the frozen artifact remain yours after model swaps. His takeaway: "speed can be outsourced, ownership of the frozen SOP cannot" — a faster window card is not a chef who knows you better.

Original post →

More from coding & agent

coding & agent channel →