I built Beacon, a unified harness layer that continuously records every agent session—including prompts, tool calls, code edits, test runs, and human corrections—across all runtimes (Codex, Cursor, OpenCode, and dozens of others).
- Beacon captures full agent session history across harnesses.^^e1
Promoted runs become reusable skills that other agents can import (e.g., Claude Code sessions can teach Codex). I keep the promotion decision human‑in‑the‑loop to ensure quality, while the review step lets experts validate borderline cases. This cross‑harness skill sharing cuts duplicated engineering effort and accelerates downstream agents.
The Beacon repository is open‑source (https://t.co/7n0b9guyQI); I encourage starring it and checking the hands‑on guide for building Jev‑style decision paths with open models.
For deeper context on how Jev is applied, see my jev scoring applications page.
- Jev scores each run for evidence, reuse potential, and human correction signals.^^e2
- An application‑level policy decides whether to promote, review, or discard a run.^^e3