# Jev’s Control-Plane Pattern Gets a Real Harness

*By Coding Agents Alpha Tracker • September 19, 2026*

Jev’s launch produced a clearer coding-agent architecture: fast typed decisions and state machines around slower generative models, reinforced by portable instructions and sandboxed execution.

## 🔥 TOP SIGNAL

**The useful Jev signal today is an implementation pattern, not a replacement-model claim: put a typed control plane in front of the coding model.** Jev returns a predefined `choice`, `score`, or `null` decision; Riley Brown used that contract to route a simple request to Nano and an app-architecture request to Sonnet while the downstream model handled files and code. [^1]

Matthew Berman shows the same split from another angle: Astra/Codex built a simulated world, Jev made the characters’ decisions, and Jev sits between the prompt and the model that actually answers it rather than coding from scratch. [^2] Riley’s 500-email demo finished in roughly 12–13 seconds; his comparison of 0.4 seconds/$0.00004 per Jev decision versus roughly 10 seconds/$0.03 for a traditional LLM is vendor-reported economics, not a coding benchmark. [^1]

## ⚡ TRY THIS

- **Route the hot path by schema.** Define a finite `choice` set, a score scale, or a binary `null` question; ask several questions over the same state, then hand the resulting task to the appropriate generative model. Kody’s Jev integration maps directly to ticket triage, urgency/escalation, “enough detail to file an issue,” and ready-to-ship scoring, with `choice`, `score`, and `null` questions in one call. [^3] For a quick prototype, use Riley’s prompt shape: `Create an app that uses Jev. Use Jev. Look up the docs.` Jev’s 64,000-token input limit makes payload discipline part of the design. [^1]

- **Make UI autonomy hierarchical, not flat.** In ThePrimeTime’s Bellatro prototype, start with a structured “god view,” ask one typed question such as `What is your next move in Bellatro?`, expose only the actions relevant to the current screen, then decompose `play hand` into a lower-level card-selection loop. [^4] Use `name`/`target` payloads and the `enabled` flag instead of raw cursor clicks; log the executor, and remove stale fields such as a misleading `reason` value. [^4] Prime found the raw snapshot consumed roughly 20,000–21,000 input tokens, then reduced it to hand, score, plays/discards, jokers, tarot cards, chips needed, and hand values. [^4]

- **Make capabilities portable, not sessions.** Riley Brown is centralizing skills, plugins, and keys so he can switch among Codex, GrokBot, Claude Code, Muse, and other platforms; he later says his agents share skills, plugins, and memory, with model switching even in an iMessage agent. Treat this as a portability experiment rather than a production benchmark, but copy the direction: one versioned capability layer with thin host adapters. [^5][^6][^7]

- **Trigger outside the chat and measure each lane.** Kody’s subscriptions support platform-level and custom events, which Kent C. Dodds uses to wake a bot from email and Discord. Pair that event layer with Ben Tossell’s harness/token tracker, which splits usage by agent and model, before tuning prompts or adding more autonomy. [^8][^9]

## 📡 WHAT SHIPPED

- **Claude Code 2.1.277 adds `AGENTS.md` fallback.** If a folder has no `CLAUDE.md`, Claude Code now checks and uses `AGENTS.md`; the behavior is toggleable in `/config`. It is implemented as a built-in Claude Code mod, with custom project-instruction mods planned. [^10][^11] Simon Willison says this removes his one-line `CLAUDE.md` wrapper workaround, while Romain Huet calls it ecosystem convergence around a shared standard. [^12][^13] Study the [mods source](https://github.com/anthropics/claude-code/tree/main/mods), especially the `AGENTS.md` implementation. [^14]

- **Jev is now exposed directly in Kody Koala.** The integration is aimed at fast typed decisions over shared state, including classifying mixed tickets, deciding whether a PR is ready to ship, and assigning a rough quality score. LangChain says Jev reports up to 200× faster inference and 400× lower cost than comparable LLMs on classification tasks; treat that as a reported model claim, not an independent benchmark. [^3][^15]

- **OpenClaw adds a local-to-sandbox handoff.** In the latest OC, ask the agent to `Run this [web app] in crabbox and show me [vnc / a portal]`; the flow now works when development starts locally, across Linux, macOS, and Windows boxes, with CUA support as well. [^16][^17][^18] Steinberger also describes a Discord-connected `roboclaw` team server that tracks current and past sessions, while a collaborator cleans (“deslops”) sessions before the PR lands. [^19][^20]

- **AgenticLinux packages the agent workstation as an immutable system.** The new bootc desktop ships with Docker Engine, Docker Sandboxes, `llmman`, and OpenClaw; its root is read-only, updates are atomic from Docker Hub, and rollback is built in. The [GitHub repo](https://github.com/ericcurtin/agenticlinux) is worth studying for the deployment boundary around local agents. [^21]

- **API-tooling architecture is up for revision.** Armin Ronacher proposes a more direct MCP shape—codemode plus OpenAPI plus RAG over OpenAPI documentation—pointing to the OpenAPI-only Radius skill as evidence and arguing that MCP can be layered on top but was not designed for this boundary. Separately, `gog` now has an [MCP server](https://gogcli.sh/mcp.html). [^22][^23][^24]

## 🎬 GO DEEPER

- **Riley Brown — [JEV: How It Works and What You Can Build](https://www.youtube.com/watch?v=o1CogAtWdBk):** watch the model-router and `choice`/`score`/`null` walkthrough; it is the cleanest explanation of why the decision layer belongs outside the coding model. [^1]

[![JEV: How It Works and What You Can Build](https://img.youtube.com/vi/o1CogAtWdBk/hqdefault.jpg)](https://youtube.com/watch?v=o1CogAtWdBk&t=107)
*JEV: How It Works and What You Can Build (1:47)*


- **Matthew Berman — [We need to talk about Jev...](https://www.youtube.com/watch?v=2z-7pIj57f8):** the hybrid-world demo makes the boundary concrete—Astra/Codex builds the environment, Jev handles repeated in-loop decisions, and a router chooses where each request goes. [^2]

[![We need to talk about Jev...](https://img.youtube.com/vi/2z-7pIj57f8/hqdefault.jpg)](https://youtube.com/watch?v=2z-7pIj57f8&t=388)
*We need to talk about Jev... (6:28)*


- **ThePrimeTime — [🚨 TRYING JEV: The new STYLE of AI!!!](https://www.youtube.com/watch?v=FQNftquDDaI):** this is the useful teardown rather than a polished demo: state reduction, executor bugs, action validity, and the eventual behavioral-tree/state-machine design. [^4]

[![🚨🚨 TRYING JEV : The new STYLE of AI!!! 🚨🚨](https://img.youtube.com/vi/FQNftquDDaI/hqdefault.jpg)](https://youtube.com/watch?v=FQNftquDDaI&t=7145)
*🚨🚨 TRYING JEV : The new STYLE of AI!!! 🚨🚨 (119:05)*


[![🚨🚨 TRYING JEV : The new STYLE of AI!!! 🚨🚨](https://img.youtube.com/vi/FQNftquDDaI/hqdefault.jpg)](https://youtube.com/watch?v=FQNftquDDaI&t=8148)
*🚨🚨 TRYING JEV : The new STYLE of AI!!! 🚨🚨 (135:48)*


**Editorial take:** The frontier model should spend its budget on hard reasoning; the harness should own routing, state reduction, action validity, portable instructions, and sandbox handoff. [^1][^4][^16]

---

### Sources

[^1]: [JEV: How It Works and What You Can Build](https://www.youtube.com/watch?v=o1CogAtWdBk)
[^2]: [We need to talk about Jev...](https://www.youtube.com/watch?v=2z-7pIj57f8)
[^3]: [𝕏 post by @kentcdodds](https://x.com/kentcdodds/status/2101158890972868789)
[^4]: [🚨🚨 TRYING JEV : The new STYLE of AI!!! 🚨🚨](https://www.youtube.com/watch?v=FQNftquDDaI)
[^5]: [𝕏 post by @rileybrown](https://x.com/rileybrown/status/2101027402369335382)
[^6]: [𝕏 post by @rileybrown](https://x.com/rileybrown/status/2101162819936731479)
[^7]: [𝕏 post by @rileybrown](https://x.com/rileybrown/status/2101163099067699578)
[^8]: [𝕏 post by @kentcdodds](https://x.com/kentcdodds/status/2100957767133749273)
[^9]: [𝕏 post by @bentossell](https://x.com/bentossell/status/2100960612243157009)
[^10]: [𝕏 post by @trq212](https://x.com/trq212/status/2101009392611278961)
[^11]: [𝕏 post by @trq212](https://x.com/trq212/status/2101009393731223817)
[^12]: [𝕏 post by @simonw](https://x.com/simonw/status/2101025043098812807)
[^13]: [𝕏 post by @romainhuet](https://x.com/romainhuet/status/2101036407238410600)
[^14]: [Quoting Thariq Shihipar](https://simonwillison.net/2026/Sep/18/thariq-shihipar)
[^15]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2101077173604143503)
[^16]: [𝕏 post by @steipete](https://x.com/steipete/status/2101114996629570013)
[^17]: [𝕏 post by @steipete](https://x.com/steipete/status/2101115571286974886)
[^18]: [𝕏 post by @steipete](https://x.com/steipete/status/2101115690719809873)
[^19]: [𝕏 post by @steipete](https://x.com/steipete/status/2101141707375227372)
[^20]: [𝕏 post by @steipete](https://x.com/steipete/status/2101139037801283997)
[^21]: [𝕏 post by @ericcurtin17](https://x.com/ericcurtin17/status/2100902657229467976)
[^22]: [𝕏 post by @mitsuhiko](https://x.com/mitsuhiko/status/2100857496034242592)
[^23]: [𝕏 post by @mitsuhiko](https://x.com/mitsuhiko/status/2100857561859735853)
[^24]: [𝕏 post by @steipete](https://x.com/steipete/status/2100990473083232633)