Jev as a Fast Decision Layer

shogo@copilot_shogo

Sep 20

This project explores the use of Jev as a rapid decision-making component that orchestrates actions around existing systems, rather than as a direct replacement for large language models.

SignalJev handles browser actions and target element selection, while a small LLM is only called when text actually needs to be generated.

I spent some time looking through what people are actually building with Jev on GitHub. The ecosystem is moving surprisingly fast, and the most interesting pattern isn't using Jev as another LLM. It's using Jev as a fast decision layer around existing systems. Here are 10 of the most popular and interesting Jev-related repos I found: ① browser-use/jev-ultrafast — ★4.8K Probably the clearest example of where Jev fits. Jev handles browser actions and target element selection, while a small LLM is only called when text actually needs to be generated. Observe → Decide → Execute → Repeat. github.com/browser-use/jev-ultrafast ② tamaratran/fast-jev-compaction — ★2.7K Uses Jev to identify unnecessary tool calls and results in Claude Code history and compact them while preserving the important context. Really interesting use of Jev inside the coding-agent harness itself. github.com/tamaratran/fast-jev-compaction ③ TheoLeeCJ/SemIf — ★1.5K An independent open-source exploration of the same general idea behind fast choice evaluation. It uses open models to directly evaluate probabilities across possible choices and can run locally. Not the official Jev model, but an interesting direction for understanding where this architecture could go. github.com/TheoLeeCJ/SemIf ④ vinnylarouge/jevlike — ★850 Another independent research project exploring how to train small models to evaluate variable sets of choices simultaneously. Again, not an open-source version of TypeSafe's Jev, but interesting research around the same decision-model paradigm. github.com/vinnylarouge/jevlike ⑤ jarrodwatts/jev-trader — ★800 An experimental trading bot using Jev to make trading decisions on Kuru on Monad. It's primarily a dry-run/mock experiment and obviously doesn't prove profitability, but it's a good example of putting Jev inside a continuous decision loop. github.com/jarrodwatts/jev-trader ⑥ TianyuCodings/NanoJev — ★310 A Qwen3-0.6B-based parallel decision model with open weights and training code. One of the more interesting attempts to explore Jev-like decision models locally. github.com/TianyuCodings/NanoJev ⑦ thruwire/foreman — ★280 Jev acts as a supervisor for Codex. Codex writes the code while Jev continuously evaluates whether the work is complete, whether verification is needed, whether the agent is stuck, or whether human intervention is required. Jev doesn't replace the coding agent — it watches it. github.com/thruwire/foreman ⑧ fhshaik/typesafe-mario — ★260 Jev playing Super Mario Bros. The emulator converts game state into structured data, Jev chooses the next action, and the controller executes it. Game state → Jev → Action. A surprisingly clean demonstration of what fast decision models are good at. github.com/fhshaik/typesafe-mario ⑨ devagrawal09/jev-review — ★250 Uses Jev to review Git diffs and repositories for correctness, safety, compatibility, and testing risks. Another good example where the model isn't being asked to generate much — it's being asked to evaluate. github.com/devagrawal09/jev-review ⑩ awlevin/typesafe-computer-use — ★200 A macOS Computer Use implementation using OCR/accessibility information to understand the screen and Jev to decide the next action. Instead of sending every screenshot through a large multimodal model, Jev becomes the fast decision layer for computer interaction. github.com/awlevin/typesafe-computer-use My biggest takeaway: Jev doesn't look like a replacement for Claude, GPT, or Gemini. It looks like something that sits between intelligence and execution. LLMs reason and generate. Jev evaluates and decides. Software executes. Browser agents, coding agents, Computer Use, trading, games, code review — once AI decisions become cheap and fast enough, you can start putting them inside loops where calling a frontier model every step wouldn't make sense. That's the part I want to experiment with next.

Original post