A weekly roundup of the Signal items we published this week (25–31 October 2026). 58 items, in the order they were published. Each links back to the original official source.
- Anthropic hands its Agent Skills format to a vendor-neutral foundation — The SKILL.md standard, already integrated in more than 70 agent platforms, moves to the Linux Foundation's Agentic AI Foundation with versioned releases planned. Anthropic (official newsroom)
- SageMaker's multi-turn RL takes a 20B model from 61% to 84% on held-out Wordle words — An AWS hands-on guide trains on whole-episode reward instead of per-step labels, and reports the gain for about $339 in training tokens. AWS Machine Learning Blog
- Coast Guard opens comment on testing super-intelligence target detection on rescue sensor feeds — A DHS notice proposes a CRADA with LOOKOUT to take automated target detection from demonstration to a measured coverage factor for search-and-rescue models. Federal Register (USCG notice)
- NIST, NSA and CISA set ten practices for incidents during model training and evaluation — The voluntary guidance covers preparation, detection, response, recovery and reporting for runs where an agent takes uninstructed actions against operator-set scope. NIST (official newsroom)
- China's State Council makes employment impact a checkpoint for AI rollout — A State Council office directive adds employment-risk review to AI project approval and asks firms to give displaced workers notice, retraining and a buffer period. State Council of China (Guobanfa [2026] No. 29)
- Alibaba Cloud's 900-run test finds a "DO NOT" prompt rule still leaks other customers' tickets — In 900 agent runs, prompt-level access rules cut cross-customer leaks but still let 10 of 12 by-ID requests through; a pre-call tool hook took that hole to 0%. Alibaba Cloud Community (AgentRun)
- Microsoft's diagnostic panel scores 83% on NEJM cases where physicians reached 36% — A Science paper pairs SDBench, a benchmark that scores step-by-step test ordering, with MAI-DxO, a multi-agent panel that debates diagnoses and weighs test cost. Microsoft AI
- Bayer's 22-agent layer answers four in five plant-breeding questions without an engineer — Bayer Crop Science put a supervisor agent over its ML workflow platform; diagnosing a failed breeding workflow fell from over an hour of log hunting to under 10 minutes. AWS Machine Learning Blog
- Updating a SageMaker endpoint can quietly leave it with no auto scaling at all — AWS documents three failure modes when endpoint configuration and auto scaling change together, including one that deploys cleanly and silently disables scaling. AWS Machine Learning Blog
- TensorRT 11.1 turns compiler knob combinations into build routes it can search — NVIDIA's Global Performance Tuner enumerates route expressions, builds every candidate, then keeps the fastest engine that still passes an accuracy check. NVIDIA Technical Blog
- Amazon Polly's generative voices get a drop-in refresh with 14 new voices — The TTS engine update adds 14 voices across 9 locales, makes Hindi and Finnish generative voices generally available, and leaves existing API calls unchanged. AWS Machine Learning Blog
- Sensetime lifts GPT-6-Astra step budget 10x, then ships a cheaper embodied stack — Sensetime reports tasks GPT-6-Astra failed 5/5 now pass 5/5 with in-context learning and up to 10x the step budget, and open-sources a 27B embodied model it says cuts cost 50x. Sensetime (official newsroom)
- Qwen-Image-2.1 folds generation, ten-image composition and alpha cutouts into one 7B model — Qwen-Image-2.1 Pro and Turbo APIs are live on Model Studio: one 7B model rendering native 2K, editing from ten references and emitting real RGBA alpha, from 0.016 dollars an image. Alibaba Cloud Community (Qwen)
- MCP async tasks never say where state lives; Alibaba puts a topic behind every task — Alibaba Cloud maps the MCP Tasks extension onto RocketMQ LiteTopic: one ordered channel per task ID, so any replica can rebuild a task state without routing affinity. Alibaba Cloud Community (Alibaba Cloud Native)
- Haiku 5.5 is the first small Claude with an effort dial, and it jumps to 72% on OSWorld — Anthropic's Haiku 5.5 lifts computer-use accuracy from 15.7% to 72.4% and Terminal-Bench from 0% to 39.2% over Haiku 4.5, at about a quarter of the running cost. Anthropic
- Nvidia's Minerva maps every symptom back to the training job that owns it — Nvidia's AIOps platform ties GPU, fabric and storage telemetry to the jobs consuming them, then correlates symptoms that each sit below any static alert threshold. NVIDIA Technical Blog
- Nvidia's self-healing engine now covers Kubernetes training, not just Slurm clusters — Mission Control 2.5 extends the autonomous recovery engine that watched HPC jobs on Slurm to Kubernetes workloads managed by Nvidia Run:ai, aiming at self-healing AI factories. NVIDIA Technical Blog
- A cached prompt only pays off if the agent stops rewriting its own prefix — AWS walks through Bedrock prompt caching for agent loops: implicit and explicit breakpoints, a 1,024-token floor, and why volatile fields must sit after the cache boundary. AWS Machine Learning Blog
- Forty production lines, forty models: at fleet scale the pipeline is the product — AWS rebuilds its edge visual-inspection pattern on SageMaker SDK v3, training an anomaly detector and a classifier per camera behind quality gates and a concurrency-limited fleet runner. AWS Machine Learning Blog
- An always-on agent is a runtime change, not a longer chat window — Alibaba Cloud's agent handbook team surveys ten personal-agent products and locates the real shift in the state kept outside the session, not in a better model. Alibaba Cloud Blog (Alibaba Cloud Native Community)
- Agents do not need more context; they need a map of the system — An open-source object-graph layer added 10 to 20 points for four flagship models on a data-agent benchmark, arguing the missing piece is system structure. Alibaba Cloud Blog (Alibaba Cloud Native Community)
- Ray jobs now resume from the record, not the beginning — EMR Serverless Spark writes record-level Ray Data checkpoints, so a failed batch job re-reads its input but skips the work already committed. Alibaba Cloud Blog (Alibaba Cloud Big Data and AI)
- Alibaba's agent teams get a control plane with a completion veto — AgentTeams 1.2.5 adds a dashboard for members, skills, tools and model routes, plus project boards that refuse to close while tasks remain open. Alibaba Cloud Blog (Alibaba Cloud Native Community)
- A refund agent that refuses its own high-confidence answers — Alibaba's case study wires a Qoder agent into a dispute queue, then has the application backend reject any decision it cannot check. Alibaba Cloud Blog (Alibaba Cloud Indonesia)
- Genie 4 puts audio and spoken dialogue inside a real-time world model — DeepMind's Genie 4 generates explorable worlds with native sound and interactive speech, steered by text while Gemini keeps the scene coherent. Google DeepMind
- CUDA 13.4 enforces a hard memory ceiling on shared GPUs, and Kubernetes gets an operator — CUDA 13.4 adds per-workload memory limits for GPU sharing, paired with an open-source Kubernetes operator that applies them from pod annotations. NVIDIA Technical Blog
- Microsoft's MAI Image 2.7 targets the edits image models usually break — Microsoft's new image editing model ranks third on Arena Edit and second on Artificial Analysis, claiming gains on text, lighting and scene edits. Microsoft AI
- Claude's drug-discovery partners report lab hits, not just designs — Anthropic published partner results: shorter CAR-T constructs, 35,000 antibody designs in a month, and a paperwork review that cut human time by 89%. Anthropic Research
- Three inference tweaks take a 124B MoE model to 54 tokens/s on one DGX Spark — NVIDIA measured a 42.5% decode throughput gain on Ling-3.0-flash from new MoE kernels, FP8 output-layer quantization and speculative decoding. NVIDIA Technical Blog
- Kimi K3.1 Pro posts 65.7% on Terminal-Bench 4.0 and promises open weights by 4 November — Moonshot's post-trained 3T-class model keeps a 1M-token context, reports 65.7% on Terminal-Bench 4.0, and adds a Highspeed variant at 2-3x output speed. Moonshot AI (Kimi official newsroom)
- Alibaba's fix for the agent handoff: split the brain, the hands and the task state — An Alibaba Cloud architecture note separates a hosted harness, cloud-held context and memory, and remote sandboxes, joined by a stable task identifier. Alibaba Cloud Native Community
- Alibaba's engineering team runs agents as digital twins, not digital employees — A practitioner account from Alibaba Cloud's Cloud Communications team wires agent stages into a pipeline with fixed input and output contracts, each agent inheriting its owner's permissions. Alibaba Cloud Native Community
- A memory layer that deletes what it cannot re-verify scores 0.93 on its own staleness test — Quen, a Qwen-powered hackathon build, treats trust as a runtime check: low-confidence memories are verified against the live source before the agent answers. Alibaba Cloud Project Hub
- Gemini 4 Carbon tops a model index at 59 and goes to cyber defenders before developers — Google's newest frontier model tops Artificial Analysis at 59, reports 88.2% on SEC-bench Pro, and ships first to vetted cyber defenders. Google DeepMind
- A vision eval where 96.8% of the labels came from the model being replaced — Property Finder found 96.8% of its vision labels came from the model it was replacing; after re-annotation, Llama 4 Maverick beat it at 86% lower cost. AWS Machine Learning Blog
- AWS runs FSDP across eight L4s with no custom container, and prices the stack at $9.50 an hour — A sample fine-tunes Qwen2.5-1.5B with FSDP over two EFA-linked g6.12xlarge nodes using the stock Ray Train container - configuration only, no image build. AWS Machine Learning Blog
- Microsoft's smaller cyber model trims 18% off a scan by taking more of the work itself — MAI Cyber 1.1 Flash matches its predecessor on Microsoft's internal MDASH scan suite at 18% lower cost per scan, and ships only inside MDASH. Microsoft AI
- Microsoft pre-trains a 2.8T-parameter MoE on 16k GB200s and publishes the utilisation — MAI-Base-2 runs 38 days at 21% MFU and 89% goodput, and Microsoft claims 4.2-12.3x better pre-training FLOPs efficiency than DeepSeek-V4-Pro-Base. Microsoft AI
- China puts agent adoption on its 2030 informatisation scoreboard — A 12-indicator plan for 2030 adds the penetration of intelligent terminals and agents, next to 6G launch timing and PUE limits for new compute. Cyberspace Administration of China (CAC)
- RocketMQ gives each agent session its own topic and holds 15 million of them at 50k TPS — Alibaba rewrote the subscription index in RocksDB so every session gets an ordered, replayable channel instead of sharing one queue with strangers. Alibaba Cloud Blog (Alibaba Cloud Native Community)
- A 4B model on the phone keeps the same clinical logic as the cloud in offline neonatal screening — LUMEN fails over between Qwen Cloud and a 4-bit Qwen3-VL-4B on device; on 150 held-out images parseable clinical answers rose from 65 to 100 percent. Alibaba Cloud Project Hub
- Nine embodied tasks flipped from five failures to five successes by raising the step limit, not retraining — SenseTime probed GPT-6-Astra across 12 simulated benchmarks and priced one pick-and-place at 263 minutes and about 55 dollars of inference. SenseTime Newsroom
- Qwen-Image-2.1 puts generation, editing and a real alpha channel behind one 7B model — A 32-layer single-stream DiT with an 8B Qwen3-VL text encoder outputs RGBA at native 2K; the Turbo tier runs 8 denoising steps at 1.6 cents an image. Alibaba Cloud Blog (Qwen)
- Claude Opus 5.5 leads Anthropic's own benchmark table at 66.4% on Terminal-Bench 4.0 — Anthropic's new flagship posts 66.4% on Terminal-Bench 4.0 and 54.4% on FrontierCode v1.1, while the company itself warns benchmark gaps are narrowing. Anthropic
- TF32x3 keeps a molecular simulation on FP32's energy surface; one-pass TF32 does not — NVIDIA shows a 1.5-1.7x faster molecular dynamics step is only worth taking when the precision trick still converges to the same geometry. NVIDIA Technical Blog
- Epsilon writes an 18-step UI test in plain English, then replays the run instead of re-asking the model — Epsilon layers Nova Act on top of Playwright: 4-6x faster test authoring, but its 100% pass rate comes from replaying a saved trajectory. AWS Machine Learning Blog
- MoEngage gives ten agents 400 private APIs without one MCP server per team — A single gateway turns 400 REST operations into permission-scoped tools, and agents find them by describing a task instead of holding every definition. AWS Machine Learning Blog
- The passport answer was wrong in perfect French - Nvidia's fix is persona-grounded procedure data — Nvidia's France and Belgium persona sets pair census-grounded residents with official procedure text, aimed at answers a fluent model gets factually wrong. NVIDIA Technical Blog
- A Qwen forensics agent whose findings are checked against artifacts before they count — A DFIR agent runs a fixed 16-step pipeline over 195 MCP forensic tools; every AI claim is validated against real artifacts, and in one public case 21 of 44 findings went to humans. Alibaba Cloud Project Hub
- Five Qwen agents argue over 139 seats and strand 3 protected passengers instead of 17 — In one frozen storm scenario a single planner stranded 17 passengers with special needs; five negotiating Qwen agents stranded 3, met every such SLA and signed each ruling. Alibaba Cloud Project Hub
- One person, 17 Qwen models and 40 dollars: a pipeline that plans every shot before rendering — Rexgent routes script, image, video and voice work across 17 Qwen models, locks character faces before rendering and scores each clip for drift, built solo on a 40-dollar credit budget. Alibaba Cloud Project Hub
- China publishes its 19th batch of deep-synthesis algorithm filings — The Cyberspace Administration published the 19th batch of domestic deep-synthesis algorithm filings on 30 October, with the October list attached and queries routed to the filing system. Cyberspace Administration of China
- Alibaba Cloud says agents created about 80% of its new PostgreSQL instances this year — Alibaba Cloud reports new ApsaraDB RDS for PostgreSQL instances grew over 400% in a year, with roughly 80% created by agents and agent traffic nearing half of all database access. Alibaba Cloud Community
- OpenAI's misalignment monitors now cover every frontier RL run, in four stages — OpenAI describes a four-stage monitoring stack for frontier agents, from activation probes to an investigative agent, plus a replay of a sandbox escape. OpenAI
- One robots.txt token is meant to carry an AI-training preference to every provider — OpenAI says it will honor a shared robots.txt signal for AI-training preferences drafted at the IETF, with implementation promised in the coming months. OpenAI
- Nvidia's agent blueprints tune the harness first and leave the weights alone — NemoClaw blueprints pair an existing agent harness with Nemotron models and OpenShell runtime controls, without rewriting the agent itself. NVIDIA Technical Blog
- Dynamo's Kubernetes recipe now pins the cluster layer under the model server — Nvidia's AICR ships validated recipes that pin the GPU driver, scheduler and gateway versions Dynamo needs, and drops NATS from the default path. NVIDIA Technical Blog
- Copyright Office renews research text-mining access and asks where the corpora go — The Office proposes renewing all but one DMCA anti-circumvention exemption, including academic text-and-data-mining access, and opens three comment rounds. Federal Register - U.S. Copyright Office