Open Source Radar — September 21: decision models go local
Today's board is the Jev ecosystem branching out. TypeSafe's decision model — typed yes/no, choice and score answers with probabilities instead of generated prose — now has a fast browser agent, a context compactor, and a trainable family anyone can run at home. Plus two harnesses worth a look.
Jev Ultrafast (Python, ~13,500 stars, MIT) — Browser Use rebuilt its agent loop around decisions instead of prose. Rather than screenshotting the page and asking a vision model what to do, it reads the visible controls into a numbered table and asks Jev one question per step: which operation (click, type, select, scroll, wait, done) and which element — two answers in a single request. A small LLM only writes text when the operation is typing. The repo's measurements put a Google Flights search from Zürich to London at 7.1 seconds end to end, with median task time down 25% and browser round-trips cut from roughly 1,090 to 101 versus its previous loop. The honest caveat is printed in the README: that's three repeats of one task on one browser profile, not a reliability benchmark. Still, it is the cleanest demonstration yet of why splitting "decide" from "generate" pays off in agent loops.
ZCode (TypeScript, ~4,700 stars, Apache-2.0) — Z.ai open-sourced its coding agent harness yesterday, and the shape is unusual: one backend serving an Electron desktop app, a browser workbench and a terminal agent, with the agent runtime and CLI source included rather than shipped as a binary blob. Provider configuration is a local file you control, so it can be pointed at whatever model endpoint you run. Two commits old and moving quickly, which is both the appeal and the risk — if you have been looking for a harness where you own the client and can inspect how the loop works, this is the freshest one on the board.
fast-jev-compaction (TypeScript, ~4,900 stars, MIT) — Claude Code's built-in compaction asks a model to summarize old turns, and summaries are lossy: the exact error string, file path or constraint can vanish even when it matters three steps later. This plugin never rewrites anything. It scores every tool call and its result with Jev in one fast request, drops or truncates the ones the model says are no longer needed, and keeps everything else verbatim. The pitch is auditability — nothing is paraphrased, a dropped call can always be re-run, and the reduction ratio tells you when pruning isn't worth it and the built-in summary should be used instead. The README is candid about the limit: a probability is not proof a result was safe to delete, and it falls back rather than failing.
Kev (Python, ~1,500 stars, Apache-2.0) — Jared Palmer's answer to "Jev, but yours": a family of small decision models at 0.8B, 4B and 9B parameters built on Qwen3.5 that answer yes/no, multiple-choice and rating questions with calibrated probabilities. Weights, training code and evaluation data are all in the repo, and the 4B and 9B versions fit a 32 GB Mac in bf16 — CUDA and Apple Silicon both supported. The API deliberately matches TypeSafe's System One, so code already written against the hosted service can be pointed at your own server. A week old and committing dozens of times a day, with per-model cards and a playground for checking whether option order changes the answer.
Agent-Native (TypeScript, ~5,500 stars, MIT) — Builder.io's framework for the part everyone gets wrong: agents that need a UI, not just a text box. You define a capability once as a typed action, the agent uses it as a tool and the UI calls the same function from code, so validation, permissions and behavior cannot drift between the two paths. The same action is exposed over HTTP, MCP, A2A and a CLI for free. It ships chat, authentication, skills and memory, scheduled automations, agent teams, and Postgres in production with an embedded database for local work. If your agent app currently has one implementation for the model and a second one for the humans, this is the deduplication.
Worth watching this week: whether typed decision models stay a Jev-shaped niche, or become the default referee layer inside every agent loop.
Would you rather an agent's context be summarized by a model, or pruned by a model that only answers keep or drop? Tell us in the comments.
Sources: Jev Ultrafast (GitHub) · Jev (Product Hunt) · ZCode (GitHub) · LocalLLaMA: ZCode is now open source · fast-jev-compaction (GitHub) · Kev (GitHub) · Agent-Native (GitHub) · Agent-Native docs
