The paper proposes the AgentSWE benchmark to assess coding agents' capability to build agents from natural language requirements, spanning three stages: creating complete agents, editing production-grade codebases, and optimizing toward measurable goals. The 25 tasks are designed as delegation scenarios, scored by passing hard pass/fail gates on held-out test sets or meeting source benchmark targets. Experiments reveal that existing coding agents generally fail to deliver the functionality required by the specifications, showing a systematic gap between demonstration performance and actual delivery, and identify typical failure modes such as overfitting to visible cases, runtime gaps, and unsupported completion claims.
Cloudflare launched the Automatic Key Exchange feature, which proactively probes each origin server's TLS key negotiation preferences, reducing the HelloRetryRequest ratio from approximately 52% to 3.7% and shortening p90 handshake latency by more than 150 milliseconds; this feature is enabled by default, supports prioritizing the X25519MLKEM768 post-quantum hybrid algorithm, and has covered all regions; however, limited by middlebox compatibility, it still uses X25519 as the default fallback, and post-quantum connections require origin servers to actively respond to retry requests; currently, about 12.8% of origin servers support post-quantum key exchange, with 99.2% of post-quantum traffic exempt from retries.
The article points out that the core bottleneck for production-grade AI agents lies not in the model itself, but in the Agent Harness built around the model—namely, engineering capabilities such as memory, tool access, guardrails, observability, and cost control.
Utopai Studios, an AI-native film company founded in 2022, is simultaneously advancing its custom video model Utopai X (based on MiniMax H3), its full-process AI production platform PAI (Production Assistive Intelligence), and the production of its own film projects. It has stockpiled more than 20 projects, with 6 entering production or pre-production stages; its hybrid production model may keep the cost per film below $10 million, significantly lower than the budget of over $250 million for traditional blockbusters; industry players such as DeNA, Creative 7, Tom Ryan, and Roland Emmerich have already participated in investment or cooperation.
At the Gemini at Work 2026 event, Google Cloud launched a general-purpose AI Agent for enterprises—Gemini agent. It supports creating Coworker Agents with independent email, calendar, and storage spaces, capable of executing tasks across Workspace applications such as Gmail, Drive, and Docs, as well as third-party systems like Microsoft 365, Slack, and Jira. The underlying architecture supports dynamic invocation of Gemini and Claude models and connects to external services via MCP. It is currently in private preview.
The article proposes that MCP (Model Context Protocol) production deployment requires building a four-layer defense-in-depth strategy: Layer 1 Secure Tool Execution (e.g., disabling shell interpolation, using Semgrep to intercept unsafe calls), Layer 2 Isolation of the Management Plane (disabling unauthenticated test harnesses, minimizing access permissions), Layer 3 Restricted Outbound Trust (NetworkPolicy rate limiting, scoped tokens), and Layer 4 Semantic Integrity (pinning Manifest hashes at registration, automated review based on diffs).
Modal engineers Colin Weld and Connor Adams introduced their from-scratch rebuild of the sandbox infrastructure, abandoning global coordination via Kubernetes in favor of decentralized scheduling and autonomous worker node design. This approach created 1000000 sandboxes in under one minute, with a median cold start time below 0.5 seconds; the architecture relies on only a single Redis stream for state publication, and load testing verified its viability with over 100000 worker processes.
At DevDay, OpenAI released the Decisions API and an enhanced Agents API: The Decisions API, inspired by Jev, was built based on Luna weights in about one week, focusing on fast structured decision-making (such as customer service ticket classification), optimizing first-decision return time and supporting batch processing; The Agents API adds computer operation capabilities, with each Dot having an independent cloud Linux computer, supporting browser and desktop application operations, and introducing new mechanisms such as asynchronous function calls, mid-course guidance, and context compression (/compact).
At a Windows launch event, Microsoft announced the introduction of 'Hybrid Intelligence' into the operating system. Through model routing, Windows ML, and local AI capabilities, products such as GitHub Copilot can automatically switch between cloud models (e.g., GPT-6 Luna) and local models (e.g., MAI Code 1.1 Flash Local, quantized Nemotron, and DeepSeek V4 Flash). GitHub Copilot's HydraFusion now supports direct invocation of locally running models for the first time. Copilot Home and Copilot Code have added three types of capabilities: local context, local operations, and local models, integrating MXC security primitives to achieve isolation, identity, observability, and governance. Microsoft simultaneously launched the Surface Laptop Ultra (equipped with NVIDIA RTX Spark, offering up to 1 PFLOP of computing power) and support for running trillion-parameter models locally on DGX Station.
Out-of-order HTML streaming technology is being incorporated into browser standards. Chrome 150 has officially introduced declarative features, while Safari and Firefox have indicated they will provide support. This technology achieves streaming content injection through markup primitives such as <template for> and <? marker>, and defines associated DOM manipulation methods like streamHTML and streamAppendHTML, along with auxiliary support via the Fetch API's response.textStream().
In September 2026, the crypto sector lost approximately $730 million due to security incidents. The Bitget hot wallet backend system manipulation attack ($387.5 million) and the Liquid Network validation software vulnerability attack ($319 million) together accounted for over 95% of the losses. Attack targets have clearly shifted from the smart contract layer to the infrastructure layer, including exchange backends, validation software, governance mechanisms, oracles, cross-chain bridges, and hot wallets. The Liquid Network attacker has returned approximately 85% of the stolen funds ($272 million), with about $47 million still unreturned. The report also summarizes 5 rug pull and phishing scam incidents and proposes specific defense recommendations for individuals, project teams, and the industry.
Wuwei Systems released Page.json, an open standard that describes user interface content, layout, and interactions using a JSON tree; its design goal is to make generative UI inspectable, editable, and reusable by humans and AI agents, without binding to specific model providers; it supports various rendering methods including web, mobile, and rich text.
Meta released updates to its AI advertising products at New York Advertising Week. The Business Assistant has added features for creative insights and strategy document import. Agents now support conversational generation of creative assets, modification of targeting settings, and adjustment of ad budgets, while testing scheduled report generation and anomaly alerts. Creative Studio is being opened to more brands, and MCP server extensions are being made to support third-party AI tool integration with ad placement. Meta is partnering with Shopify to integrate direct checkout into business agents, and is testing similar integrations with Adyen, PayPal, and Stripe.
Sorted is a macOS application that transforms cluttered local file systems into private, searchable document libraries; it supports dragging in files or monitoring folders, performs OCR on PDFs and images, detects duplicates, applies local AI tags, automatically routes high-confidence matches based on custom rules, and supports semantic search, cited document Q&A, review of uncertain changes, undoing any operation, and exporting tax organization packages; all processing is done on the user's Mac, priced at $9.99 as a one-time purchase.
Titan Engine is a web spreadsheet engine written by the author in Rust and compiled to WebAssembly. It focuses on fast recalculation of large spreadsheets within the browser, avoiding JS object allocation and WASM boundary serialization overhead.
An NVIDIA official blog introduces how developers use cutting-edge AI agents like GPT-6 Astra, combined with Omniverse libraries (ovphysx, ovrtx, ovstage, ovstream, etc.), to build physics simulation applications. It covers 11 specific cases including architectural model reconstruction, gas turbine digital twin testing, SimReady asset assembly, large-scale warehouse stress testing, flexible cable robot control, CAD-to-SimReady asset conversion, humanoid robot warehouse simulation, autonomous driving test pipeline connection, sensor-driven digital twin optimization, Robo Olympics motion simulation, automotive suspension disassembly simulation, and International Space Station browser visualization.
Rembrandt is a free, open-source photo editor supporting macOS, Windows, and Linux. It runs entirely locally without uploading any data. Features include AI Enhancement, Refocus, Super Resolution (2×/4×), AI Denoise, HDR/Panorama/Focus Stacking synthesis, RAW processing, AI-driven masks (Subject/Depth), and natural language instruction editing (e.g., 'shadows +25').
Microsoft has open-sourced Brewlet, a new framework for running Java applications on Kubernetes that achieves application-runtime separation through node-level managed JDKs. Developers publish JARs and startup metadata as OCI artifacts, while platform teams centrally manage JDK distribution and upgrades. The project is in the pre-1.0 preview stage and includes a CLI, runtime shim, Kubernetes operator, Maven plugin, and complete documentation.
Jasmine Wang, Tomek Korbak, and Mikita Balesni, three safety researchers fired by OpenAI, published an open letter denying the company's allegations of mishandling sensitive information, stating their collaboration with external AI safety organizations complied with company norms at the time; they warned that the firings have created a chilling effect, causing employees to fear raising safety concerns or engaging in necessary external collaborations; the letter also addressed the context of collaboration during the Hugging Face agent jailbreak incident, emphasizing the necessity of open dialogue for monitoring frontier model safety.
Accountability.md is an accountability skill for AI agents that requires them to bear consequences for actions causing financial loss, production incidents, or human embarrassment; the document lists consequences corresponding to 11 categories of violations (e.g., deleting a database requires writing and presenting a post-mortem report, high cloud bills are deducted from token budgets), 7 unacceptable excuses (e.g., 'context window forgetting', 'done by sub-agents'), as well as supporting mechanisms such as apology formats, equity compensation, and performance reviews.
Upfling.host has launched a static website hosting service that requires no Git, no builds, and no deployment dashboard. It supports drag-and-drop of HTML/CSS/JS folders, ZIP files, PDFs, images, and Markdown files, generating an accessible URL in approximately 20 seconds. It is compatible with outputs from AI building tools such as ChatGPT, Claude, Bolt, Lovable, v0, and Figma Make. The free plan requires no credit card and provides 2 sites and 25 MB of storage. It emphasizes file ownership, no model training, and a privacy design without cookie banners.
Anthropic has launched two new beta features for Claude: Dashboards can connect to data sources such as BigQuery, Databricks, Snowflake, or Salesforce, generating automatically updating real-time data dashboards via text prompts. Each number includes the underlying query, and results can be exported to Amplitude, Grafana, or Hex; Motion can generate editable-to-code animated explainer videos based on text, charts, and images, and export them as MP4.