#部署/工程

Smart Approvals in AWS Quick: Human oversight where it matters

AWS describes Smart Approvals in AWS Quick, which uses a separate classifier model to review eligible agent actions against the user's request and either approve them automatically or return them for human review, without granting additional permissions or removing required approvals. Policy controls such as Custom Permissions, Quick Guardrails, connector action permissions and user settings determine eligibility, and each automatic approval is single-use and applies only to the exact action and parameters reviewed. The post also covers CloudTrail and CloudWatch Logs audit evidence and recommends piloting Smart Approvals with a defined workflow before wider rollout.

Speeding Up eBPF File Opens by 54% (Not AI Gen)

The author describes three small changes to an open-source eBPF agent (between commits 9c521bd and cfc7538): skipping the LPM cache lookup when a path tail is shorter than 128 bytes, checking the inode cache before building the file path in lsm_file_open, and loading only lsm_inode_permission, tp_exit_openat and tp_exit_openat2 when ptrace blocking is enabled. In a benchmark opening one file 200,000 times and 20,000 new files, repeated opens dropped from 4.68 µs to 2.15 µs (about 54%) and cold opens from 6.33 µs to 5.36 µs (about 15%), with repeated-open p99 falling from 8.3 µs to 3.46 µs. The author notes he runs the perf test after each commit and gives the results to an LLM to find inefficiencies to fix.

AI Data Centers Turn to Decommissioned Jet Engines as Grid Queues Stretch Years

Facing grid connection waits of about seven years in places like Northern Virginia, AI companies including xAI and Crusoe are buying converted jet engines from suppliers such as ProEnergy to generate power on-site in roughly 24 months. The units are less fuel-efficient and more polluting than purpose-built turbines, and state permits for xAI's Mississippi installation allow higher emissions limits. Analysts describe the approach as a short-term fix, with Google opting to wait for grid-connected power instead.

Introducing Knowledge for Mastra Agents

Mastra introduces Knowledge, a shared store of markdown pages organized into org, resource, and thread scopes that agents can search, read, and write through five built-in tools, working alongside signals to keep context current. Agents receive a Knowledge instance via the knowledge option, and storage adds mastra_knowledge_* tables, with KNOWLEDGE_ROUTES enabling a Studio page view. Knowledge is in alpha and requires @mastra/[email protected] or later.

ElevenAgents integrates natively with leading contact center platforms to deliver enterprise-ready agents

ElevenAgents now integrates natively with NiCE CXone, Genesys Cloud, Amazon Connect, and Five9, and connects over SIP to any other contact center or PBX that supports SIP trunking. The voice stack, built from ElevenLabs' own audio research, offers more than 17,000 voices in over 90 languages, passes call context from the contact center to the agent, and hands calls back to human agents through existing routing with full context. The company says this lets enterprises automate routine calls without replacing their telephony, routing, compliance recording, or CRM systems.

vLLM Delivers Day-0 Support for Kimi K3.1 Pro

vLLM announces day-0 support for Kimi K3.1 Pro, Moonshot AI's 2.8T-parameter (104B active) multimodal MoE with a 1M-token context that adds a DSA sparse-attention indexer to all 24 full-attention layers of its KDA + MLA hybrid. The post details a packed KV cache for three kinds of state, token-sharded prefill attention (up to 1.69× faster TTFT at 340K tokens on 8× GB300), decode context parallelism with an exact top-k merge, a fused low-latency decode path, a DSpark speculator preview, AgentX benchmark results on GB300 and Rubin, zero-mismatch RL, and support on NVIDIA, AMD, TPU and Alibaba T-HEAD Zhenwu PPUs.

How U Mobile built agentic telco operations with Amazon Bedrock AgentCore

U Mobile's Information Systems Division, working with AWS and AWS Partner Axrail, built AiDaH, an AI-Driven Agentic Helpdesk on Amazon Bedrock AgentCore, which now handles over 150 documented operational scenarios across six business divisions. According to U Mobile's internal measurements, monthly trouble ticket volumes fell by up to 84.5 percent in provisioning and network tickets, 81.6 percent in general inquiries, and 48.7 percent in billing and charging between January and September 2026. The post details AiDaH's six-layer architecture, its security foundation including a Well-Architected Security Review of the Landing Zone, and the spec-driven development approach using Kiro.

Use NVIDIA NeMo to Turn Agent Failures into Tests and Harness Fixes

NVIDIA describes using NeMo to convert production evidence into prioritized issues and custom evaluations for improving agent prompts, tools, models, memory, skills, and harness code. The post notes that generic benchmark scores may not explain domain agent failures, such as choosing the wrong tool, following the wrong workflow branch, missing relevant context, or using an unnecessarily expensive model.

Reviewing AI-Generated Code

The article argues that reviewing AI-generated code carries all the cost of human code review with little benefit, since the AI won't change its behavior based on review feedback and experienced reviewers' skills will atrophy over time. It suggests that having AI review AI code is no more effective, and that teams should instead invest in custom static analysis tools and prompt refinement, reserving human approval for change requests rather than line-by-line code.

Human-in-the-loop authorization for AI agents

Airlock is a human-in-the-loop authorization system built on the HARP specification that intercepts sensitive AI agent commands, encrypts them with AES-256-GCM, and routes them to a mobile approver who signs a decision with Ed25519 that is verified locally before execution. It features a zero-knowledge gateway, replay protection via nonces and expiry timestamps, and enterprise deployment within an organization's own infrastructure. The HARP protocol is described as an open, tool-agnostic standard covering canonicalization, artifact hashing, decision signing, and replay protection.

Show HN: AI Gude - Personal Podcast News Aggregator

AI Gude is free, self-hosted open source software that turns the RSS feeds you follow into a personal morning podcast: at scheduled times it reads new articles, writes a script in your style, speaks it with an AI voice and delivers one audio episode with chapters. It uses Gemini for text and speech, Firecrawl for article text, ffmpeg for encoding and SQLite for storage, and supports multiple users via a reverse proxy that sets the X-Aigude-User header. A live demo is available at aigude.alexdoes.dev with the user and password "demo".

Daybreak Red is now available in AWS GovCloud (US-East)

Daybreak Red brings GPT-5.6 Cyber to approved customers in AWS GovCloud (US-East) through Amazon Bedrock for authorized cybersecurity work.

See the agents, set the terms: New visibility and control for personal agents

Cloudflare launched new Security Analytics features that surface personal agent traffic, including a Personal Assistants category in the Bot Analysis chart, a Top statistics widget listing visiting agents, and per-user agent visibility for sites using Account Abuse Protection. It also expanded the Agents policy to allow per-agent settings of Allow, Block on pages with ads, or Block, renaming the setting to AI Crawler and Agent access policies. Cloudflare says its next goal is helping sites preserve user context when agents act as intermediaries.

Together Compiler: Continual Compilation for AI inference

Together AI introduces Together Compiler (TKCC), a continual optimization loop that uses high-throughput agents to generate inference-engine optimizations, built on "persistent engines" that hot-reload candidate forward passes without restarts and a two-tier engine design balancing measurement throughput against fidelity. TKCC reports up to 2.6x higher TPS across frontier open-source models in under 24 hours, state-of-the-art low-latency results on SemiAnalysis' InferenceX benchmark (1111 TPS on MiniMax M3.1-Flash, 1205 TPS on DeepSeek v4.1-Flash), and 1.7x over open-source megakernels on Llama 1B at batch size 1, enabled by 252x higher experiment throughput than baseline agentic systems. The post also says TKCC generalizes to weather modeling, PDE, time-series, and biology workloads, and is serving trillions of tokens today.

Show HN: Id|le, little AI-creatures living in the browser (story mode)

A Show HN project presents "Id|le," little AI-creatures living in the browser in a story mode, where messages arrive at an inbox and each stage is a classifier creature built from a real text classifier from Jeffy. The three starting classifiers (spam, support tickets, mood) run in the browser after a one-time 23.7 MB download, while any other chosen head is queried on Jeffy's server. A filter drops matches into its own bin, a router sorts into a bin per label and sends uncertain items to review, and a tagger stamps a label and passes it on, with one message every 4 seconds.

When AI Fails: Building Fault-Tolerant AI Systems

Aman Sardana argues that AI components should be treated as dependencies that can fail, and that classic reliability patterns—timeouts, circuit breakers, bulkheads, retries, fallbacks, and graceful degradation—must be extended to cover AI's quiet failures, where a model returns a fast, well-formed, confident but wrong answer. He proposes a semantic circuit breaker that trips on validation failures and low confidence, layered response validation, a deliberately designed fallback chain, and a pre-agreed degradation table specifying fail-open or fail-closed behavior for each use case.

How to make scientific data archives AI-ready

The article argues that scientific data archives are poorly suited to AI use because data is locked in opaque files, and proposes a four-step architecture: place data in scalable object storage, index chunks rather than filenames, co-locate dataset-level metadata with the data in object storage, and catalog repositories rather than individual files. It says this approach lets agents and humans quickly find and fetch only the chunks a query needs, citing a 576x reduction in data fetched for a GOES-16 timeseries and a 100x speedup in common NASA queries. The piece describes this as how Earthmover's Arraylake works, built on Icechunk and Zarr, and notes the same recommendations apply to humans and AI alike.

Sovereign AI is how Europe could catch up

This promotional page for Aleph Alpha's Kolibri presents it as a sovereign AI model for the public sector, available via API, with the company offering to build specialized sovereign language models tailored to an organization's context. It cites customer results including an AI assistant for an agency serving 80,000 users, AI agents extracting verifiable insights from sensitive documents at a global chipmaker that cut search times by 90%, and AI support in requirements engineering at an automotive supplier that led to 40% faster RFQ processing. The page also lists the company's management team.

MCP vs. API vs. CLI: How should AI agents connect to your app?

This post argues that an MCP server should be the first choice for making a web app agent-friendly, because MCP standardizes both server and client behavior, lets the client handle OAuth so the model never sees the token, and can route destructive-action confirmations directly to the human via elicitation. REST APIs remain better for scripts, bulk jobs, data exports and integrations that run without an agent, while CLIs suit agents with a shell working with local files. The author recommends putting business logic in operations and exposing them as thin adapters over MCP, REST and optionally a CLI, using the open-source Linkly link shortener as the example.

Snowflake's AI_PARSE_DOCUMENT got an agentic mode

A practitioner tested Snowflake's new AGENTIC mode for AI_PARSE_DOCUMENT on a SharePoint-fed Cortex Search pipeline, finding it costs a flat 8.18 credits per 1,000 pages versus 0.68 for OCR (about 12x) and only paid off on slide decks, where it returned chart, diagram, table and infographic summaries absent from OCR text. Two undocumented behaviors surfaced: image summaries live in each page's images array rather than the page content, and the mode parameter must be a literal, since a CASE expression silently returns NULL. The author settled on AGENTIC for PPTX and OCR for everything else, appending only the four useful image types to page text.

Polymorphic Inference: self-tuning inference for open models

Together AI introduces "Polymorphic Inference," an approach for adapting inference to fast-changing model architectures, traffic shapes, and hardware, contrasted with the older "Static Inference" paradigm. The company announces Together Compiler, an AI-agent-driven "Continuous Compilation" optimization loop already running in production on some serverless endpoints, and previews upcoming work including self-serve compilation, shape-shifting speculative decoding, the ThunderAgent Router, post-training, and TogetherLink workload-aware model routing.

Together Compiler: Continuous Compilation for AI inference

Together AI introduces Together Compiler (TKCC), a continual agent-driven system that recompiles inference engines as traffic drifts, using persistent engines, a two-tier fast/slow engine setup for throughput-fidelity balance, and a custom cost model for parallelism selection. TKCC reports 1.3-3x higher TPS across frontier open-source models in under 24 hours, state-of-the-art low-latency results of 1111 TPS on MiniMax M3.1-Flash and 1205 TPS on DeepSeek v4.1-Flash on SemiAnalysis' InferenceX benchmark, and 1.7x over open-source megakernels on Llama 1B batch size 1, enabled by 252x higher experiment throughput than baseline agentic systems. It also generalizes to weather modeling, PDE, time-series, and biology workloads.

Featured coverage of this story: 《Together Compiler: Continual Compilation for AI inference》

How we built Lovable, and why we keep rebuilding it

Lovable's founder recounts how the platform evolved from the 2023 GPT Engineer project into a product where users have created more than 70 million projects, repeatedly reinventing its agent design—from complex async subagents back to a predictable tool chain, then to "knowledge files" with retrieval instead of fine-tuning. He describes stateful sandboxes that keep development servers running for faster edit-check cycles and a framework called Camille for forking and merging entire running computers. He closes with three open problems: automating the full software lifecycle, making all types of software seamless to build, and building a better software stack.

EdgeDAE: Acceleration of Diffusion Action Experts for Real-Time Physical AI

EdgeDAE is a heterogeneous FPGA-GPU system that offloads the perception-heavy vision transformer to the GPU while accelerating Diffusion Action Expert (DAE) inference on an FPGA by storing all parameters on-chip in BRAM/URAM, eliminating the repeated DRAM parameter loading of iterative denoising. It co-designs quantization, fixed-point arithmetic, and hardware-efficient random number generation for the FPGA fabric. Compared with an edge GPU baseline, it reduces end-to-end inference latency by 52.5% for Octo-Small and 38.7% for Octo-Base with up to 2.10x higher throughput, and achieves about 17x higher energy efficiency than an RTX 4090.

Browser Use vs Kernel: Pricing, Stealth & Performance

This comparison contrasts Browser Use and Kernel cloud browser providers across pricing, stealth, performance, authentication, and developer experience. Browser Use offers headless browser time at $0.02/hour with no subscription, while Kernel bills per second at roughly $0.06/hour headless and $0.48/hour headful, with plans from $0 to $200 per month. The article notes Browser Use reached 90.5% of protected sites versus Kernel's 80.5% in a Browser Use-published study, and that lifecycle benchmark totals were 371.5 ms for Browser Use and 398 ms for Kernel on October 27.

Webflow wanted $2,500 for one language. AI rebuilt our whole site in a weekend

Rezi.ai describes moving its ~515-page site off Webflow to an Astro 7, React 19 and Tailwind CSS 4 stack hosted on Netlify, with Claude Opus 5.5 doing the heavy lifting, after Webflow quoted $2,500 a year to add a single language on top of a roughly $30,000-per-year enterprise contract. The team first tested the approach on two smaller brands, then kept the Rezi.ai rebuild visually identical using a script that compares every page to the Webflow original, and reports desktop PageSpeed rising from 38 to 99 and mobile from 57 to 95.

Show HN: Meteosource MCP server, weather data for AI assistants

Meteosource released an MCP server over Streamable HTTP at https://mcp.meteosource.com/mcp that authenticates with a Meteosource API key sent in an Authorization: Bearer header (or X-API-Key), with no OAuth or sign-in. It exposes seven tools: search_places, get_forecast, get_alerts, get_weather_history and get_air_quality on weather plans, plus get_solar_production and get_wind_production on the Renewables plan. Setup guides for Claude Code, Claude Desktop, claude.ai, Codex, Antigravity and other clients are in the README's Connect section.

GPT-Synopsys Combines IC Design EDA with Agentic AI

At its 2026 Investor Day in New York, Synopsys announced GPT-Synopsys, a specialized AI model built with OpenAI's frontier models that operates directly in Synopsys EDA tools for chip design and verification tasks including PPA optimization, timing closure, and verification. Running on OpenAI-hosted infrastructure and integrated with Synopsys.ai and the Synopsys Autopilot agentic AI platform, it will be trained and tuned specifically for chip design, with OpenAI licensing Synopsys' EDA tools and paying a training subscription fee, while the two companies share revenue based on how well the model improves chip design. Synopsys says proprietary IC design data will not be used to train the frontier model and will be protected through access controls, encryption, and runtime guardrails.

I built a free local AI co-processor in Rust. AXON Developer Edition

AXON Developer Edition is a free, local AI co-processor built in Rust that runs on Windows 10/11 and lets AI assistants access files, memory and tasks on the user's own machine without sending data to the cloud. It is not an AI model itself, and users connect assistants such as Claude via its MCP connector or other web-based AIs through AXON LOCKA, a temporary authenticated tunnel. The full build requires no account and is downloaded as a ready-to-run binary that listens on port 8880.

A private AI box in one apply

Conn3x now has circuits for Ollama and Qdrant, and its Ollama repo includes an example that builds a private retrieval-augmented generation box on a DigitalOcean droplet with a single conn3x apply. The setup runs Ollama and Qdrant on loopback only, with SSH as the sole open port, and the post walks through storing a note, querying it by meaning, and having qwen2.5:0.5b answer. The author reports 101 resources applied in 2 minutes 36 seconds, a clean second plan, and a 41-second destroy.

"Start with a Monolith" Was Good Advice. AI Is Changing That

A Medium post from Scalar Engineering's "Architecture in the AI Era" series (Part 1) revisits the common advice to start with a monolith and argues that AI is changing that calculus. The submission links to the article and its Hacker News discussion thread.

Local AI app that runs an 80B model on an 8GB Mac

A local AI app called Yanah claims to run an 80B-parameter model on an 8GB Mac at chat speed, keeping data private with no cloud, subscriptions, or internet needed to chat. The post explains that models store learned patterns as parameters in a file rather than the internet, and that a capable AI can fit in under 50 gigabytes.

v2.1.318

Claude Code v2.1.318 adds an answer record when Claude Code asks to enable mods, assume_role on mantle upstreams of the Claude apps gateway, an effort option for a mod's $.model.fork, and sign-in expiry dates in claude auth status (loginExpiresAt) and claude -p JSON results (credential_expires_at). It fixes numerous issues, including --continue losing earlier conversation after malformed tool call retries, macOS stalls with many running Bash commands, permission prompts answered by stray keystrokes, and config save failures on shared network home directories. It also improves auto-update retries, MCP tool errors, remote session reconnects, and adds a No repository option for cloud sessions.

Adamanteum: A local, MIT knowledge atlas your notes and any AI can both read

Adamanteum Community Edition is a local-first, MIT-licensed knowledge atlas that renders a Markdown vault as an interactive map with fast full-text search, a notebook with revisions, and a code graph for 30 languages. It runs via three commands on Node.js 24 or newer, keeps notes read-only by default, and lets any MCP-capable AI search and read the vault through revocable read-only tokens, with optional OpenAI-compatible model connections. The post reports one-laptop benchmarks, such as indexing 10,000 notes in 10.4 seconds, and lists limitations including incomplete Obsidian parity for editing, tags and embeds.

CrowdStrike Introduces the AI Kill Switch for AI Safety

CrowdStrike announced the AI Kill Switch in Falcon Guardian, which lets security teams terminate a rogue AI agent where it runs, block it from relaunching on the targeted host or host group, and cut off its enterprise network access, with an authorized administrator triggering the action and every action recorded in the audit trail. The capability builds on the Falcon sensor and on what Falcon Guardian already sees—what an agent is told to do, what it communicates, and what it executes—and is available today.

Soniox STT v6 Adds Agent Text Input and Echo Filtering for Voice Agents

Soniox released STT v6, a speech-to-text model available as stt-rt-v6 for streaming and stt-async-v6 for files across 60+ languages, which accepts the voice agent's own text to better recognize short caller replies and filter out leaked agent audio. Agent text input is enabled via an "enable_agent_text_input" flag on the real-time model only, with text streamed as JSON frames over the same WebSocket as caller audio. Soniox has published no v6 accuracy or latency numbers; v5 scored 1.34% semantic WER on the Pipecat benchmark, and v5 models are deprecated with removal on December 15, 2026.

NVIDIA Shows 3 Ways to Speed Up MoE Inference on DGX Spark, Lifting Ling-3.0-flash Throughput 42.5%

NVIDIA's technical blog describes three stacked optimizations for running Ant Group inclusionAI's 124B-parameter Ling-3.0-flash (5.1B active) on a single DGX Spark at batch size one: Humming MXFP4 MoE kernels, an FP8 LM head, and built-in NEXTN speculative decoding. Across all 1,319 GSM8K questions, decode throughput rose 42.5% (37.81 to 53.89 tokens/s) and mean latency fell 24.5%, with accuracy essentially flat (96.89% to 96.82%). NVIDIA notes that combining the FP8 LM head with MTP requires a shared LM-head module, otherwise the framework silently falls back and throughput drops below the unoptimized baseline.

Data centers are turning to old jet engines as the wait for power grows

Some AI data centers are using converted jet engines rebuilt to burn natural gas to generate their own electricity, as grid connections and new turbines face years-long waits, the New York Times reported. Rystad Energy analyst Audun Martinsen said jet-powered units can have a data center running in 24 months, compared with about seven years to hook up 100 megawatts to the grid in Northern Virginia, according to the IEA. xAI has installed eight converted jet-engine turbines in Mississippi for its Colossus 2 data center in Memphis, and Crusoe ordered 13 of ProEnergy's converted units in April for OpenAI's Stargate site in Abilene, Texas.

Lotse: AI Coding Agent – iOS

Lotse is an iOS app that connects to selected GitHub repositories and lets you describe a change on your phone; the agent then works in the cloud, runs tests and opens a pull request for review, sending a push notification when finished. It supports tests for JavaScript/TypeScript, Python, Go, Rust, Java/Kotlin and Flutter/Dart, and offers a free tier plus Pro subscriptions ($4.99 monthly or $24.99 yearly). The app, formerly Jules Console, notes that Google ends Jules on December 31, 2026, and that task descriptions and repository code are sent to AI model providers via OpenRouter with user consent.

Fireworks unveils new products at Forge to help companies build specialized intelligence

At its first annual Forge conference, Fireworks launched products aimed at helping companies build AI around their own data and tasks, including Ultra-fast inference on NVIDIA Blackwell GPUs reaching up to 1,500 tokens per second, a half-price Background Mode, and multi-day prompt caching. It also introduced Fireworks Nexus for bringing open models into existing coding and coworking tools, No Eval and Specialized Model Inference for training and serving custom models, and Fireworks Sovereign, now in private preview, which runs its inference engine inside customers' own cloud or data centers. Fireworks says it already manages over 25,000 GPUs in customer environments.