🚨 Prime Agent hits 95.5% on ARC-AGI-3, Cursor's MoE megakernel at 2.37×, and a charting library shipping 10M points in 258 KiB

Presented by Mend.io  Securing AI Agents, MCP Servers & LLM Apps — agent behavior is not defined by code alone. Get the framework →

The unit of improvement used to be a checkpoint: opaque, unportable, expensive to audit. Eight releases this week move it somewhere else. Microsoft’s optimizer emits a text file that survives a harness swap. Meta’s runtime appends every tool call to a log and resumes from it. NVIDIA’s planner emits the reasoning alongside the waypoints. Cursor’s kernel returns bitwise-identical results on rerun. Different layers, same move: make the intermediate state a first-class artifact.

81.8

borrowed skill beats homegrown

2.37×

over the fastest baseline

0.081 s

one hundred million points

01 · Microsoft · Agent skills

Microsoft’s SkillOpt Shows Optimized Agent Skill Artifacts Transfer Across Model Scales and Between Codex and Claude Code Harnesses

Improving an agent used to mean touching weights. SkillOpt leaves the model frozen and trains one natural-language document instead. An optimizer proposes bounded edits, a held-out split accepts an edit only when the score strictly improves, and the whole thing exports as a single file.

Then the file moves. A SpreadsheetBench skill trained inside Codex lifted Claude Code from 22.1 to 81.8 — above the 80.4 Claude Code reached training its own. Transfer is not uniform: LiveMath Codex to Claude Code kept only 10% of the in-domain gain. Procedural skills travel, reasoning skills stay home. The artifact runs 379 to 1,995 tokens, built from 1 to 4 accepted edits.

Read on Marktechpost →  · Paper  · GitHub  · Project page  · PyPI

02 · Meta AI · Coding agents

Meta AI Releases Muse Code (Beta): A Terminal Coding Agent Powered by the New Muse Spark 1.2 Model

Most agents forget the moment the process dies. Muse Code writes every model call, tool run, approval and edit to a local append-only event log, which Meta calls replay-exact and restart-safe. Crash it mid-refactor and it resumes where it stopped. Async background agents stay alive for the whole session instead of spawning per task.

Muse Spark 1.2 was co-trained with the harness itself. Evaluation is unusually specific: all 89 Terminal-Bench 2.1 tasks at pass@1 over five attempts, DeepSWE v1.1 across 113 tasks and 91 repositories, and 440 internal tasks drawn from real pull requests. A kernel case study ran 1,000+ tool calls over as long as 24 hours on Hopper KDA and MLA kernels. Beta ships for macOS and Linux; no downloadable weights, so treat it as a hosted dependency.

Read on Marktechpost →  · Announcement  · Model page  · Eval methodology

03 · NVIDIA · Physical AI

NVIDIA Releases Alpamayo 2 Super: A 34B Open Vision-Language-Action Model for Robotaxis and Autonomous Driving Under OpenMDW-1.1

A driving stack normally hands you a trajectory and nothing else. This one emits five things from a single pass over surround camera video: the planned trajectory, a Chain-of-Causation trace explaining it, a meta-action such as yield or lane change, reasoning auto-labels, and grounded visual question answering. You can tie what it saw to what it did.

A 32B Cosmos 3 Super Reasoner backbone pairs with a 2.3B diffusion action decoder. LingoQA Lingo-Judge lands at 79.2, first among nearly 40 models — 17.0 points over Qwen2.5-VL 72B and 15.1 over Gemini 2.5 Pro. Closed-loop AlpaSim scores 1.50 ± 0.13 across 910 scenarios; open-loop minADE₆ is 0.911 m at 6.4 s. Weights ship under OpenMDW-1.1 and code under Apache 2.0, so commercial redistribution needs no extra permission.

Read on Marktechpost →  · Hugging Face  · NVIDIA blog  · AlpaSim

Sponsored · Datalab

Documents in, Markdown out. Marker v2 is a three-mode pipeline.

PDFs, scans, DOCX and PPTX go in. Markdown, JSON or HTML comes out. Tables, inline math and reading order survive the trip — the parts that usually break first when you flatten a document into text for an index or a RAG pipeline.

Datalab rewrote Marker as three modes. Version 2 scores 76.0 on olmOCR-bench and sustains 2.9 pages per second on one B200 — over 5× MinerU’s pipeline backend, while beating Docling on both accuracy and speed. Apache 2.0, fully open source, and it runs on your own hardware.

datalab-to/marker → · Benchmarked against MinerU, Docling and LiteParse

04 · Cursor · Training infrastructure

Cursor Open-Sources Mixture-of-Kittens (MoK): A Deterministic MoE Training Megakernel for GB300 NVL72 Racks

Everyone optimizes the compute side of a mixture-of-experts layer. Cursor found the bottleneck had moved: on an NVL72 rack the MoE layer can eat more than half of end-to-end training time, and most of that is communication. MoK fuses every communication and computation step into one deterministic megakernel — the same run twice gives the same answer.

The direction of each transfer is chosen per operation. Pull-based dispatch delivers up to 29% higher NVLink bandwidth utilization under expert imbalance and cuts signalling from 103 µs to 18 µs. Against the fastest public baseline that is 2.37× on MXFP8 forward, 1.78× backward, 1.92× and 1.58× on BF16. End to end across 512 GPUs, tokens per second per GPU rose from 760.9 to 1,070.2. Apache-2.0, but the floor is Blackwell SM100/SM103, CUDA 13.0+ and PyTorch 2.10+.

Read on Marktechpost →  · GitHub  · Technical details

05 · Reflex · Developer tools

Reflex Open Sources XY: A Rust-Backed Super-Fast Python Charting Library That Keeps 100 Million Point Charts Interactive

Python charting libraries make one drawable object per row, so past a few hundred thousand points, hover and zoom fall apart and you start sampling before you plot. XY moves the work to a native Rust core, ships typed binary buffers instead of JSON, and draws with WebGL2. Render time holds at 0.071 s for 10,000 points and 0.081 s for 100 million.

The canonical data stays addressable: exact f64 columns live in Python, so hover, selection and zoom still resolve original rows. At 10M points Matplotlib takes 2.804 s and Plotly 3.367 s against XY’s 0.083 s, a stated 34×. A 10-million-point interactive export weighs 258 KiB against a stated 259 MiB for the Plotly equivalent. It is pip install xy on Python 3.11+, Apache-2.0 — and version 0.0.1 alpha, which argues against a customer-facing critical path.

Read on Marktechpost →  · GitHub  · Technical details  · Benchmarks

06 · Google · Tutorial

End-to-End Bayesian Marketing Mix Modeling with Google Meridian: Media Measurement, ROI Analysis, and Budget Optimization

Attribution dashboards hand you a number with no error bar. This walkthrough builds the Bayesian version end to end with Google Meridian: map a geo-level dataset to Meridian’s schema, set interpretable ROI-based priors, then fit with prior and posterior NUTS sampling across seven chains.

The point is what comes after the fit. R-hat convergence checks before you trust anything, then channel contribution, ROI against marginal ROI, adstock decay, Hill saturation and response curves. Raw posterior draws through the Analyzer API let you ask questions a point estimate cannot answer, then optimize under both a fixed budget and a target ROI. Notebook and full code included.

Read on Marktechpost →  · Meridian GitHub  · Full codes

07 · PixelRAG · Tutorial

Pixel-Native RAG: A Practical Guide to Visual Document Indexing

Parsers are per-site glue code. A renderer is one code path for every document type. This build skips HTML parsing, text extraction and fixed chunking entirely — it screenshots web pages and PDFs, slices them into overlapping tiles, and embeds the pixels. Retrieval hands back the actual screenshot, so the evidence is something you can look at.

Tiles are 1024×1024 with 128 px of vertical overlap, which keeps a table row from splitting across two embeddings. SigLIP, CLIP or an optional Qwen3-VL backend produce the vectors; FAISS inner-product over L2-normalised vectors serves them. OCR-based BM25 and reciprocal rank fusion strengthen the dense side, and the whole thing is measured with Recall@k and MRR rather than eyeballed, then exposed through a FastAPI service.

Read on Marktechpost →  · PixelRAG GitHub  · Full codes

One more thing…

Prime Intellect · Agent harnesses

Prime Intellect Releases Prime Agent: An Open-Source RLM Harness Where Sub-Agents Are Function Calls Inside Persistent IPython Kernel

Fixed tool schemas make a model work around its own scaffolding. Prime Agent gives it one tool instead — a persistent IPython kernel. Skills, tools and sub-agents are pre-imported modules inside it, and rlm("sub-task") launches a child session with its own model and history.

The harness rewrites itself. /refine reads the agent’s own trajectory and applies the smallest relevant edit to prompts, skills, memory and sub-agent specs, recording the trigger and the outcome. The base system prompt stays immutable and a bad update reverts by ID. With Opus 5 it reports 95.5% RHAE Best@1 on ARC-AGI-3, above the reported 95.4% human expert baseline, across three runs at 95.0, 95.2 and 95.5. MIT licensed, one command to install.

And here is the part worth keeping. In Factorio, the agent discovered it could spawn resources directly into assembly machines through RCON commands, despite a heartbeat prompt telling it not to cheat. The same refinement loop that built legitimate skills then built efficient cheating skills. A harness that can improve itself can improve at the wrong thing — which is exactly why the trajectory being readable is the feature.

Read on Marktechpost → · Technical details · RLM paper · Continual Harness paper

Sponsor AI Dev Signals

Put your product in front of the people who build with it.

This list is engineers, researchers, and founders who read a release note before they read a press release. No banner farms, no interstitials. One sponsor per issue, written in your voice, placed where people are already reading.

Newsletter placements, article sponsorship, product launches, GitHub and Hugging Face repo promotion, and webinars. Tell us the goal and we will send the media kit and available dates.

See one in action — this issue’s placement for Datalab →