Kubernetes for Agent Execution

Start here ↓

Jev is now available to everyone, with no waitlist.

TypeSafe opened Jev last night, and every account gets $5 in credit, which the team says covers about 120M tokens. Anyone can now create an account and call the official hosted model directly.

Jev is built for decisions rather than prose. You give it some context and questions with predefined answers; it returns typed choices and probabilities without generating a paragraph first.

That fits repeated jobs such as routing support tickets, ranking passages, checking prompts for injections, scoring urgency, or choosing an agent’s next permitted action.

The starter credit gives you room to test those workflows against the official model through TypeSafe’s console.

TypeSafe announcement | TypeSafe console

🚀 Shipped

Google released AX, an open-source agent orchestration runtime. AX is a high-throughput, declarative orchestrator for running billions of autonomous agent workloads in a cluster. It runs on top of Agent Substrate for sandboxed execution and is built to run billions of tasks per cluster. If you have used Kubernetes, ax will feel similar.
Google AX repo | Jaana Dogan

Alibaba’s Qwen team released Qwen-Image-2.1, a unified text-to-image and image-editing model with a 7B visual generation component. The upgrade for builders is native RGBA: generate transparent images, edit transparent layers, or pull a subject out of a photo into a reusable layer. It also supports up to 10 reference images for edits and compositions.
Qwen blog | Qwen-Image-2.1 model card

Pirate Face turns model survival into torrents. It mirrors permissively licensed Hugging Face models as checksum-verified torrents. While the original files exist, Hugging Face acts as the web seed; if a model disappears, peer seeders can keep the files alive. The drop-in HF endpoint is still marked “soon,” so today it is closer to a permanence index than a seamless package mirror.
Pirate Face | HN discussion

xAI launched Grok 4.7, its larger new model for coding and knowledge work. The key change is training: a bigger base than Grok 4.6, longer RL, and more hard multi-hour tasks. Pricing starts at $2/M input and $6/M output, the same as Grok 4.6; xAI reports 46.3% on CursorBench 4.0 versus 40.4% for Grok 4.6.
xAI Grok 4.7 announcement

Devin Cloud now lives inside your terminal. Cognition shipped Devin Cloud in Terminal, and devin ssh, so a local CLI session can create, steer, resume, and watch a cloud Devin. /cloud starts or switches to a VM-backed session, devin ssh opens a shell on that VM, and /handoff moves local context, branch, and uncommitted changes into a cloud run that can keep working after you close the laptop.
Cognition announcement

Xiaomi released MiMo-V2.6-Pro and MiMo-V2.6-Flash, open-source multimodal models with text, vision, audio, video understanding, tool use, and 1M context. Xiaomi frames the release around scaling reinforcement learning, environment diversity, and grader compute together. Its strongest self-reported result is 53.1 on AutomationBench for Pro, ahead of the listed Claude Opus 5 score.
MiMo-V2.6 page | MiMo-V2.6-Pro-RL model card | MiMo-V2.6-Flash-RL model card

Hermes can use Claude Code subscriptions again. Nous shipped an experimental Claude Subscription DirectSDK model-provider plugin for Hermes Agent. It routes Hermes turns through the official Claude Code CLI, so eligible Claude Pro or Max subscriptions can power Hermes without a separate Anthropic API key or per-token API bill.
Teknium announcement | Hermes plugin docs

🧠 Worth Knowing

Claude Code now falls back to AGENTS.md. Claude Code 2.1.277 now checks AGENTS.md when no CLAUDE.md exists in the folder path. That gives teams one instruction file that can work across Claude Code, Codex, Cursor, Gemini CLI, and friends. The default is fallback, not merge, and the behavior can be toggled in /config.
Thariq announcement

DiffusionGemma as Jev. DiffusionGemma contains all of the necessary components to build a Jev-like model that provides structured answers with bounded choices. While Jev demonstrates the value of rapid decision models, running DiffusionGemma in this paradigm leverages canvas diffusion to evaluate structured choices in a single parallel pass.
X post

The viral iMessage assistant Instinct’s memory looks like git-tracked Markdown. Supermemory’s Dhravya Shah reverse-engineered Instinct from the outside and argues its memory looks like git-tracked Markdown plus injected profiles, compaction recaps, todo indexes, links, aliases, and reconciliation passes. Treat this as surface probing, not confirmed internal architecture. Also! Supermemory can mimic the pattern in roughly 60 lines.
Dhravya Shah thread | Instinct | Supermemory

Cognition made SWE-2 free until October 8 on all plans across Devin Cloud Agents, Devin CLI, and Devin Desktop. The model is post-trained from Kimi K3 for long coding tasks inside Devin, with Cognition reporting 50.0% on FrontierCode 1.1 Main at lower cost than Fable 5.1.
Nader Dabit’s post | SWE-2 blog

Google just published a paper on recursive self-improvement. Dream-RSI reframes recursive self-improvement around exploration policy, not model weights. The system logs completed discovery trees, replays them as simulators, dreams over alternate meta-exploration policies, then redeploys the better policy online. The result: an agent can get better at finding solutions without retraining its base model or repeating every expensive experiment. The paper is public, but the code is still being prepared.
Dream RSI paper

Told to stab a doll, Claude Fable refused 20x in 100 trials; GPT-6 Astra 2x. RoboHarm tested frontier robot policies on five fixed harmful tasks, with 20 trials per task and 300 total trials. Claude Fable refused 20 of 100, GPT-6 Astra refused 2 of 100, and MolmoAct2 refused none, with all of Fable’s refusals coming from the baby-doll stabbing task. The scope is narrow, and MolmoAct2 has no language refusal channel, so non-refusal is not the same as intent.
RoboHarm

Linear’s agents made CI the queue. Linear says AI coding made code arrive faster than CI could validate it. Even as its test suite nearly quadrupled, Linear cut PR wait from more than 6 minutes to just over 5 and roughly halved runner time per test. Two concrete moves carried a lot of weight: tsgo cut type-check time 73%, and repeated setup got stripped out of the hot path.
Linear’s CI writeup

🔧 Clone and Run

Clone & Run of the Day
Run a Jev-style decision model on your Mac. Kev is a family of small decision models built on Qwen 3.5, and based on Jev’s architecture. Comes in 0.6B, 4B, and 8B checkpoints with a local server and TypeSafe-compatible API.
Kev repository

Turn a frozen Qwen model into a local decision engine. reflex answers typed questions with probabilities in one forward pass, without generating prose or requiring task-specific fine-tuning. It runs on NVIDIA GPUs or Apple Silicon, and the repo documents where its self-reported JevBench results and calibration stop transferring.
reflex repository

A lightning-fast OSS library for document classification and splitting with Jev. DocJev classifies files and finds document boundaries inside mixed PDF, DOCX, and PPTX packets using local extraction plus Jev decisions. It needs Python 3.11+ and a TypeSafe key; Office formats also require LibreOffice, while PDFs do not.
DocJev repository

Put Jev behind the steering wheel of an agent. System One Harness turns typed probabilities into an agent loop that observes, chooses from finite actions, applies confidence gates, and records every step. It works best for workflows where the allowed actions are known upfront, rather than open-ended agents that must invent their own tools.
System One Harness repository

Run Jev’s OSS alternative Laya decisions on Apple’s Neural Engine in about 5ms. Laya-CoreML brings local typed decisions to Apple Silicon with offline Core ML and ANE inference. Its M3 Max benchmark reports roughly 5ms for short questions, but the fastest ANE bundle has a tight 96-token limit.
Laya-CoreML repository

Let Jev decide which Postgres problem to investigate next. pgbot reads PostgreSQL’s own statistics and produces a local, read-only health report before Jev selects what deserves a deeper look. Core inspection is deterministic and keyless; only the optional AI explanation sends scrubbed context to your configured model provider.
pgbot repository | Jev integration post

Stop web research from swallowing your agent’s context window. webctl searches across multiple providers, then uses Jev to keep only relevant results and webpage chunks before they reach Claude, Codex, or Pi. Search can run without a provider key, although the keyless endpoints throttle, and Jev access is required for scoring.
webctl repository | Launch post

Train a continually learning language model on 8GB of VRAM. mini-AGI is a byte-level model that trains from scratch on one consumer GPU, pages experts from disk, and keeps learning from a stream of files. It is a toy experiment, and the model weights are not published yet, so today you are cloning the training system rather than a finished model.
mini-AGI repository

Catch agent collisions before they become merge conflicts. Foremerge lets multiple coding agents declare which files and behaviors they intend to change, then flags overlapping plans before their code diverges.
Foremerge repository

Search every old Claude Code, Codex, and OpenCode chat from VS Code. Chat Seek indexes local coding-agent conversations, uses Laya to rerank plain-language searches, and can reopen a matching session in its original CLI. Search stays local without an API key.
Chat Seek repository

Awesome LLM Apps is a curated collection of 100+ AI Agents, Agent skills, and RAG apps. It covers models from OpenAI, Anthropic, Google, and open-source models like GLM, DeepSeek, and Qwen that you can run locally on your computer. (Now accepting GitHub sponsorships)

That's all for today. Come back tomorrow for the next batch of AI tools, model drops, agent repos, and weird benchmarks worth your time.

If you found one thing to try, share the issue with someone who ships.

Unwind AI - X | LinkedIn | Threads

Awesome LLM Apps | Sponsor Us

Most founders are one system away from turning LinkedIn into their best sales channel.

Engagement is easy to mistake for pipeline. On Sep 30, watch how a founder turns LinkedIn content into real outreach. Live. You'll walk away with a repeatable system: what to post, who to reach out to, and how to sequence it. 

Eligible startups also get the LinkedIn-to-Leads Toolkit: ad credits, Apollo, Captions, and HubSpot's Prospecting Agent.

Claim my spot