#部署/工程

How TCS built an agentic test orchestrator with Amazon Bedrock AgentCore

TCS built a multi-tenant Agentic Test Orchestrator on AWS that turns Jira stories, BRDs, Swagger specs, and GitLab sources into test scenarios, Gherkin files, and executable Java or Python tests, using specialized agents hosted on Amazon Bedrock AgentCore runtime with the Strands Agents SDK. The platform separates a shared control plane from per-tenant application planes with isolated Cognito user pools, KMS keys, S3 buckets, and DynamoDB tables, and routes API tests to Amazon ECS on Fargate and Selenium UI tests to an EC2-backed ECS cluster. TCS reports that in an internal benchmark of about 100 scenarios, average effort fell from roughly two hours to about 14 minutes per scenario, an 88 percent reduction.

Introducing Security Factory

Factory introduces Security Factory, now in Private Preview, which connects investigation, remediation, and verification into one continuous workflow to help engineering teams turn security findings into tested fixes. It combines Connectors for consolidating findings and context, Workstreams for parallel agent investigations, and Factory's coding workflows for developing and testing fixes. Organizations can configure agent autonomy levels, from presenting remediation plans for human approval to automatically creating pull requests and deploying fixes within approved policies.

Boxes

Amp introduces boxes: always-on computers starting at $18 per 30 days (billed by the minute) to which users can deploy the software they build with agents, with each application getting its own URL and no containers involved. Boxes can be shared with a workspace, can be turned on and off on a schedule, and each comes with its own runner so agents can operate directly on the box.

How to Build a Voice AI Agent

This tutorial builds a Python restaurant-booking voice agent that chains speech-to-text, an LLM with tool calling, and text-to-speech through a single OpenRouter API key, streaming each completed sentence to TTS instead of waiting for the full reply. Across 12 turns in three test conversations, median server-side time to first audio was 1.9 seconds (1.4 seconds without tools, 2.4 seconds with a tool call), and a four-turn conversation cost about $0.013 per minute of processed speech at October 8 promotional prices, with TTS accounting for roughly 90% of model cost. It also compares STT, LLM and TTS models on accuracy, latency and price, and shows how to run the same agent with LiveKit Agents for live audio and interruptions.

Best OCR Model for Scanned Documents, Measured

The author tested 15 vision models plus Mistral OCR and Tesseract on eight synthetic scanned pages scored by character error rate and 264 key-value checks, finding Claude Opus 5.5 most accurate (262 of 264 at $16.42 per 1,000 pages) and Mistral OCR the best value (258 of 264 at $2.02). The post documents failure modes such as fabricated card digits, wifi passwords and server names on a faded receipt, and provides code for transcription, disabling reasoning, cross-checking two models from different labs, and using the mistral-ocr engine.

How We Built TLX: A Hardware-Native GPU Compiler for Production ML at Scale

Meta introduces Triton Low-Level Extensions (TLX), an extension to Triton that adds explicit warp-group and cluster-level orchestration so engineers can use modern GPU features like TMA, TMEM, and warp specialization while keeping Triton's programming model. TLX is in production across Meta workloads, with reported gains including 20-25% end-to-end MFU for GEM model training on thousands of NVIDIA Blackwell GPUs, up to two-thirds net-latency reduction in RecSys inference, and a distributed grouped GEMM kernel running in large-scale LLM pretraining. The post frames TLX as implementing a "MIMW" abstraction between CUDA's SIMT model and Triton's block-level model, and covers five extension areas: warp-group orchestration, synchronization, memory orchestration, cluster execution, and optional layouts.

Optimizing Muse Spark Open on TPU v7 with SGLang-TorchTPU

The post describes ExecuTorch's new WebGPU backend, which lets PyTorch models exported once to a .pte file via torch.export run in any WebGPU-enabled browser through a WebAssembly runtime, covering LLMs, vision-language, segmentation, depth, detection, super-resolution, speech-to-text and on-device LoRA/DiReFT fine-tuning. On an Apple M1 Pro in Chrome, WebGPU was 4–7× faster than ExecuTorch's Wasm/XNNPACK CPU path on prefill and 2–10× faster on decode, with all data staying on the device. The backend is still in early development, and the post links an ExecuTorch WebGPU Playground and Hugging Face repos with .pte files and export recipes.

Featured coverage of this story: 《ExecuTorch WebGPU: PyTorch Models on the GPU, in the Browser》

Introducing Mastra FDE

Mastra announced Mastra FDE, a forward-deployed engineering program open to any team building agents, in which a dedicated Mastra engineer embeds with the customer's team from architecture through production at 6 hours/week consultative or 20-40 hours/week hands-on. The post cites Salesforce's Agentforce Vibes harness, MongoDB's Sage platform, and Workday moving security operations agents from Bedrock Agents to Mastra in 31 days while cutting 3,000 lines of code and about 98% of estimated vector storage costs. It positions the offering as cheaper than one senior AI hire and backed by a shared Slack channel with Mastra's founders and engineers, roadmap insights, upskilling sessions, and Mastra Enterprise.

NVIDIA Expands Open Source Commitment With OpenShell Donation and New AI Infrastructure

At KubeCon + CloudNativeCon, NVIDIA is contributing its OpenShell agent security runtime to the CNCF as a Sandbox project and contributing to Google-led Agent Substrate, also accepted into the CNCF Sandbox. NVIDIA also announced new DSX AI Factory Platform components: NVIDIA Kubernetes Engine, NVIDIA Kubernetes Recipes and NVIDIA Snapshot, which restores checkpointed replicas in seconds without model reload or warmup. Design partners Crusoe and Saturn Cloud reported reduced workload readiness and vLLM startup times using Snapshot.

Running AV Reasoning Models Within Onboard Compute and Latency Limits

AV developers need reasoning models that fit onboard compute and latency constraints; larger models provide a foundation, but deploying them in a vehicle requires reducing model size, preserving useful reasoning, and validating performance on target hardware. The source mentions NVIDIA Alpamayo 2 Edge in this context, but the supplied material is truncated before detailing it.

Decision Models in Hugging Face

Hugging Face explains decision models, which take a state plus typed questions (choice, score, and noul, a Bernoulli yes/no primitive) and return structured decisions with probability distributions instead of generating text. The post contrasts them with traditional and zero-shot classification, and describes ecosystem support: Transformers' DecisionPipeline, AutoModelForDecision and DecisionConfig, a /v1/systemone endpoint also served by llama.cpp, a Hub tag for discovery, and fine-tuning via the Trainer API.

How to make flash models even faster: optimizing DeepSeek V4.1 Flash 1017

Baseten describes using an internal harness to optimize 80+ kernels and bottlenecks for DeepSeek V4.1 Flash 1017, producing a Fast variant it reports at 991 output tokens/s on Artificial Analysis with the lowest time per task. The post details three hardware-level techniques: kernel fusion (including an improved fused router kernel cutting decode time by 2.7ms and a 30% faster mHC megakernel), Programmatic Dependent Launch to overlap kernel execution, and host loop minimization via breakable CUDA graphs in SGLang.

Manage Kubernetes Clusters for GPU Infrastructure with NVIDIA Kubernetes Engine

NVIDIA announced two open source projects, NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR), for building and operating managed Kubernetes services on GPU infrastructure. Both are new projects available in public preview. The announcement notes that AI Factory customers often want to consume GPU capacity through the Kubernetes APIs, tools, and ecosystem they already use.

Start and Restore GPU Workloads in Seconds with NVIDIA Snapshot

NVIDIA Snapshot is a new open-source component of the NVIDIA DSX AI Factory Platform that makes initialized GPU workloads reusable across inference replicas, speeding startup and recovery. The post notes that when inference demand spikes or a replica fails, available GPUs do not immediately translate into serving capacity, since initializing an LLM workload—a cold start—can take minutes.

Agent Substrate: Advancing Kubernetes for Agents at Scale

AI agents spend most of their time idle, waiting on a human prompt, a tool result or a model response, and Kubernetes was not built for that. Running one Pod per session holds memory for inactive sessions, while deleting Pods between turns incurs a multi-second cold start and loses the agent's state, and RL post-training hits the same wall. Agent Substrate has now been donated to the foundation named in the source.

Google Cloud launches Gemini for healthcare.

Google Cloud announced Gemini for healthcare, an AI solution designed to unify fragmented clinical, operational, and administrative systems by turning siloed data into coordinated action. The platform integrates specialized skills and data connectors with existing third-party systems, drawing on electronic health records, billing platforms, clinical guidelines, and Medicare datasets to orchestrate multi-step workflows such as medical records synthesis and prior authorization review within defined guardrails and with human oversight. It is initially available in preview with launch partners including HCA Healthcare, Highmark Health, and Northwell Health.

v2.1.322

This release adds a fast mode on Vertex in the global, us and eu regions, falling back to standard speed if Vertex refuses it. The release is immutable, so only the release title and notes can be modified.

How skills and agents work in AWS Quick

AWS Quick uses configurable AI agents that follow instructions, load reusable skills, act through connectors, and ground answers in spaces, knowledge bases, datasets, and dashboards. Skills follow the Agent Skills specification as folders with a SKILL.md file, and can be built from chat, published, and shared with teams. The post also covers best practices for production skills and the AWS Quick official catalog, including the Skill Architect skill.

rust-v0.169.0-alpha.8

This is a GitHub Changelog entry titled "rust-v0.169.0-alpha.8." The supplied material consists only of the site's navigation menu and product links, with no details about the release itself.

JPMorganChase expands Claude Code to more employees across the firm

JPMorganChase is expanding Claude beyond engineering to product teams, advisors and other employees, with infrastructure and operations teams using Claude Code for monitoring, alerting and root cause analysis and the security team using frontier models to find and fix vulnerabilities faster. Claude Code runs in controlled environments including the firm's Devspace platform, with staged access and an approval process for new models, and an automated gate enforces a hard pass or halt so Claude cannot sign off on its own work. In Asset & Wealth Management, small teams are using Claude Code in their engineering workflow, and J.P. Morgan Asset Management recently announced a connector bringing 55ip's ActiveTax Technology into Claude for Financial Advisors.

What is an API endpoint? A guide to connected software

An API endpoint is the specific URL where an API receives requests and returns responses, with each route pairing an HTTP method with a resource path. The guide covers how endpoints process requests, REST methods like GET and POST, seven common API architectures, design best practices, testing and monitoring, and security through authentication, scoped permissions and concurrency-based rate limits, illustrated with ElevenAPI examples.

Autofill Passwords with 1Password and Browser Use

This tutorial explains how to connect 1Password to Browser Use Cloud via a read-only service account so an agent can fill passwords and TOTP codes, using a dedicated vault and a public test site (The Internet) to verify a login. It covers creating the vault and service account, connecting the token in Browser Use, running the login task through the dashboard or the V4 REST API, and troubleshooting common failures.

Control Room: A central system to track and stop model workloads across our research infrastructure

OpenAI describes its Control Room, a central system that tracks model workloads across its research infrastructure and can rapidly stop them when safety or security concerns arise.

Announcing Long-Context Training on TPUs with MaxText Context Parallelism

Google Cloud announced long-context training support in MaxText, its open-source JAX training framework, via four context-parallel strategies on Cloud TPU: All-gather, A2A (DeepSpeed Ulysses), P2P (Ring Attention), and the A2A + P2P hybrid (USP), all using the Tokamax Splash Attention kernel. Benchmarks report Llama 3.1 8B reaching 70% MFU on TPU 5p for sequences up to 4 million tokens, with context extendable to 14 million tokens on TPU 5p and TPU 7, where 14M-token P2P training on 64 TPU 7 chips reached 48.7% MFU versus 37.1% on TPU 5p. The feature is available in MaxText v0.2.4 on PyPI, Docker images and GitHub main, with sample recipes for TPU 7 and TPU 5p.

The best Python agent harnesses in 2026

This roundup compares Python agent harness SDKs — libraries that supply file and shell tools, planning, sub-agents, context management and sandboxes around a model — judging each on model portability, how much can be taken apart, where commands run, approvals, durability, long-run context handling, tracing and commercial support. Pydantic AI Harness is named the overall pick, with LangChain Deep Agents, the Claude Agent SDK, OpenAI Agents SDK, Codex SDK, OpenHands Software Agent SDK, Strands harness and Microsoft Agent Framework Harness each recommended for specific stacks. The piece states its figures were read from each project's own docs and repository on October 30, 2026, with the Pydantic AI Harness and OpenHands Index figures re-read on November 9, 2026.

NVIDIA Nemotron 3.5 Super VL is now available in Amazon SageMaker JumpStart

NVIDIA Nemotron 3.5 Super VL, a 120B-parameter vision-language model with 12B active parameters, a C-RADIO vision encoder, and up to 1M-token context, is now deployable in Amazon SageMaker JumpStart via SageMaker Studio or the SageMaker Python SDK. JumpStart offers an NVFP4 variant on G7 instances and BF16 reference weights on ml.p5en.48xlarge, with measured throughput of 63.07 tokens/sec on ml.g7.24xlarge and 140.85 tokens/sec for BF16 on ml.p5en.48xlarge. The model targets document intelligence, GUI understanding, video question answering, and multimodal agentic applications.

Single-shot and agentic retrieval with Amazon Bedrock Managed Knowledge Base

Amazon Bedrock Managed Knowledge Base now supports single-shot RAG and retrieval-only requests through AgenticRetrieveStream, alongside existing agentic retrieval, letting one API serve simple and complex questions. The maxAgentIteration range is now 0–5: 0 returns passages with no model calls, 1 performs one retrieval plus one generated answer, and 2–5 run up to n minus 1 planning rounds before answering. The post illustrates each mode with a fictional banking policy assistant and advises choosing the lowest iteration limit that meets each workflow's quality needs.

Make a Robot Arm Feel Weightless with NVIDIA Holoscan 5

NVIDIA describes using Holoscan 5 to build a gravity-compensation loop that lets a robot arm be guided by hand, with its motors supplying just enough torque to support its weight and keep it upright. The post says this interaction is a useful starting point for robots that learn by demonstration, manipulate objects, or work alongside people, and that such a loop requires continuously calculating the inverse dynamics.

Manage your telemetry in Logfire, send it anywhere

Pydantic Logfire has launched Fleet Management in Beta on every plan, letting users change what their OpenTelemetry Collectors and application SDKs collect and where they send it from Logfire, without editing config files or redeploying apps. Clients connect over OpAMP and pick up published, versioned policies within seconds, covering settings such as sampling, minimum log level, attribute redaction and scrubbing across all exporters the SDK exports through. Supported targets include host Collectors, Kubernetes clusters via the OpenTelemetry Operator, existing Collectors through the opamp extension, and Python, Node.js, Java, .NET and Go applications.

Professional work deserves professional tools. Meet AWS Quick.

AWS announced its first brand campaign for AWS Quick, an AI workspace built only for work where teams use AI agents to research, analyze, and act across the data and applications they already rely on. The campaign runs in the United States and the United Kingdom on streaming TV, online video, social media, and leading business publications, and cites enterprises including 3M, BMW, Mondelez, New York Life, and Southwest Airlines as users. AWS says Quick has no consumer version and integrates with security and compliance tools such as Microsoft Purview, AWS IAM, and Microsoft Active Directory, and carries certifications including HIPAA, FedRAMP, SOC 2, and ISO 27001.

How CrossnoKaye grounds physical AI with Amazon Bedrock AgentCore

CrossnoKaye's ATLAS software-defined automation platform and its AURA agentic system run on AWS, using Amazon Bedrock AgentCore code interpreter microVMs to execute model-authored Python for verifiable statistics and physical balances across plant data. AURA operates in three modes—interactive Chats with warm sandbox sessions, repeatable Tasks, and automated Routines—and enforces layered safety through Bedrock Guardrails, read-only air-gapped sandboxes, deterministic edge base control, and independent hardware protection. The post describes the architecture and reports results such as investigations completing in under four minutes and a 34 percent reduction in batch drying-time standard deviation.

Pluggable Agentic AI Framework: A seven-layer composable enterprise agentic AI reference architecture on Amazon Bedrock AgentCore

AWS presents a seven-layer composable reference architecture for multi-agent AI applications—Data and Knowledge, Model and Inference, Orchestration, Security and Guardrails, Application, Observability, and Agentic Operations—each with a distinct responsibility and stable contract, illustrated by a retail customer support reference implementation using a supervisor agent delegating to Order and Refund specialists. The post argues that pluggability at layer boundaries lets organizations swap models, tools, and providers without redesigning the stack, and that security, governance, observability, and continuous evaluation should be designed in from day one rather than added after the system is built.

The little things that shipped in October

ComfyUI's October update (v0.44) added a Branch Node for selecting one of up to 32 wired paths, native Load LoRA (Model) and Load LoRA (Text Encoder) stack nodes supporting up to 20 LoRAs, and Comfy Desktop templates that automatically download their example inputs and required models. It also brought zip downloads for multiple outputs, EXR/HDR previews and a Convert SDR to HDR node, ProRes output in Save Video, and new tools to hide, box-select and delete links, plus per-widget Undo and Redo.

Turn buying signals into pipeline with Sendoso gifts in Genie One

Databricks describes how Sendoso's MCP server, installed from Databricks Marketplace as a governed MCP Service in Unity Catalog, lets Genie One users find intent signals in their own data and then select, preview, approve, track or cancel a gift without leaving the workspace. Send history sits alongside pipeline data for attribution, while Unity Catalog permissions, Unity Gateway controls and Sendoso budgets and approvals govern each send. The post cites Sendoso's 2026 Digital Fatigue Report finding that 84% of buyers are fatigued by cold calls and 83% by marketing emails.

Extending LLDB to Debug Applications That Span CPUs and GPUs

Meta describes a vendor-neutral accelerator plugin architecture in lldb-server that separates shared debugger infrastructure from hardware-specific device control, representing the CPU and GPU as separate LLDB targets within one debugging session. The plugin translates a vendor's device debugging API into LLDB concepts such as processes, threads, registers, memory, modules, and stop reasons, and is activated only when the application uses a supported GPU runtime. Most of the common framework has been upstreamed to LLVM, including the plugin interface, GDB-remote packets, and hardware-independent mock tests.

Portals on Your Runner

Portals, previously available in orbs, now work on local machines, letting you ask Amp to start a dev server or build a website and open it from the thread's Portal tab regardless of location. Enable them via Allow Portals under Settings › Runner in the macOS app or by starting your runner with amp --no-tui --remote-control-portals; portals on your runner are accessible only to you. To share them with your workspace, start a shared runner with --share or turn on Share With Workspace in the macOS app.

ElevenAgents Campaign Manager works every contact to an outcome

ElevenLabs introduces ElevenAgents Campaign Manager, which takes a list of contacts and a target outcome and runs outbound campaigns, following up over time and carrying context between calls until the outcome is reached or the contact asks to stop. It offers campaign analytics with per-contact decision timelines, a needs-attention queue, CSV export, and controls including calling windows, attempt limits, do-not-call lists and disclosures, and it uses existing ElevenAgents agents and phone numbers over phone calls, WhatsApp voice and WhatsApp messaging, with email and SMS coming soon.

Accelerate GPU training on Unity Catalog tables with UCDataset

Databricks released UCDataset in Beta on AI Runtime, a native PyTorch data loader that streams Unity Catalog tables directly to every GPU in a training job without exports, copies, or custom loader code. On a simulated production workload streaming a 300 GB feature table across 8 NVIDIA H100 GPUs to train a recommendation model, UCDataset cut training time by 10x on average compared with loading data directly from Parquet. The Beta reads Delta tables in Unity Catalog, applies filters and deletion vectors as data is read, and enforces Unity Catalog permissions, though it cannot read tables with row filters or column masks.

Your Experts Have Partners. There's HOPE for Keeping Them Together

This post introduces HOPE (Higher-Order Pruning of Experts), a second-order expert-pruning method for Mixture-of-Experts language models that accounts for pairwise expert interactions via an F-matrix and a quadratic program, rather than scoring each expert independently. Evaluated on Qwen3.5-122B, Qwen3.5-35B and GLM-4.5-Air with two calibration datasets, HOPE achieved the best average rank across conditions and gains of up to +6.1% on agentic coding tasks over the next-best method, with its advantage growing at aggressive pruning rates such as 50%. The post also describes a three-step calibrate/solve/prune package and notes limitations, including weaker advantages at 10–20% pruning and no cross-layer interaction modeling.

What is Pathways? Single-controller training for large models

Google announces the General Availability of Pathways on Google Cloud Platform, describing the distributed system it built to train large models including Gemini. The post explains where SPMD execution falls short—pipeline bubbles, RL loops with distinct programs, and failures that stop every process—and how Pathways instead has a single Python client declare the whole computation as a graph of blocks and tensors, with a distributed runtime placing it on virtual devices, moving and resharding tensors, ordering work across the cluster, and remapping around a failed host. It also contrasts Pathways with Ray, framing Ray as orchestration and Pathways as the execution layer beneath it.