## Weekly Overview - **Anthropic completes its “model → platform → infrastructure” evolution**: Claude Code launches `/ultraplan`, `Routines`, and `Managed Agents`—elevating the coding assistant into an event-driven, cloud-hosted, composable agent infrastructure layer. - **The agent ecosystem hits a “Harness standardization” inflection point**: Agent Harness is now officially recognized as the first stable abstraction layer. Platform-level implementations—including EverOS, Vercel Open Agents, and Claude Managed Agents—are shipping simultaneously, making cross-framework reuse a reality. - **The capability gap between U.S. and Chinese large models has effectively closed**: Stanford’s *2026 AI Index Report* confirms parity on key benchmarks like reasoning and multimodality. Domestic models are accelerating in vertical domains: JD’s JoyAI-Image-Edit (spatial intelligence), Alibaba’s Qwen3.6-A3B (MoE-based coding), and Baidu’s Wenxin NabuOCR (ancient character deciphering) all match world-class performance. - **An AI-native computing paradigm is solidifying**: Claude Code deeply integrates Browser Use (infinite cloud browser), Chrome DevTools MCP (native frontend debugging), and Cloudflare Wrangler (a command-line hub for 3,000+ APIs)—unifying terminal, browser, and cloud services into a single programmable compute substrate. - **Benchmarking faces a systemic trust crisis**: Berkeley’s BenchJack experiments and ClawMark’s multi-day collaboration benchmark jointly expose environmental hijacking vulnerabilities in mainstream leaderboards like SWE-bench. Top-performing models score only ~55% on extended, multimodal tasks—revealing a clear capability ceiling. - **Hardware agents are nearing mass production**: BrainCo’s Revo 3 dexterous hand (22 DoF + tactile feedback), Geely’s i-HEV powertrain (48.41% thermal efficiency + AI energy management), and MOVA’s V70 Ultra (16 cm extendable robotic arm) have all broken through critical physical-world interaction bottlenecks. ## Hot Topics 1. **Claude Code officially launches `/ultraplan`** https://www.bestblogs.dev/status/2042850992149221732 **What it is**: The first solution to close the loop between *cloud-based intelligent planning* and *one-click local execution*. It decomposes complex dev tasks—like refactoring microservices or deploying CI/CD—into verifiable substeps. After orchestrating logic and analyzing dependencies in the cloud, it generates safe, executable local scripts—dramatically reducing cognitive load and error rates. — **Try this**: Run `/ultraplan migrate-to-turbopack` in your Next.js project. Observe whether it auto-generates `turbo.json`, dependency update commands, and rollback scripts. Track execution time and manual interventions—and compare efficiency gains against traditional migration. 2. **Agent Harness is formally recognized as the first stable abstraction layer for AI agents** https://www.bestblogs.dev/status/2042612328701812789 **What it is**: A milestone signaling the shift from “stitching model calls together” to *modular engineering*. Harness standardizes and encapsulates tool registration, context management, error recovery, and observability—enabling skill modules and execution protocols to be reused across models (Claude, Gemma, Qwen, etc.). — **Try this**: Refactor one of your existing Slack bots using Vercel Open Agents. Extract its core functionality (e.g., meeting note generation) into a standalone Harness module, then inject it into a new project via `harness.register()`. Validate whether it works seamlessly across applications. 3. EverMind’s Open-Source All-in-One Agent Platform EverOS and the Neutral Benchmark EvoAgentBench https://www.bestblogs.dev/status/2044054552639627375 Core idea: Provides open-source infrastructure covering the full agent lifecycle—creation, testing, and evaluation. EvoAgentBench is the first neutral benchmark focused specifically on *multi-day collaboration* and *cross-modal state consistency*, directly addressing core weaknesses in today’s agents: fragmentation, memory loss, and cross-modal contradictions during long-horizon tasks. — Practical tip: Plug your agent into EvoAgentBench’s `multi-day-email-thread` test scenario. Run it for 3 rounds and check whether it can still accurately reference financial data from an attachment sent on Day 1—when queried on Day 5. If it fails, prioritize reviewing your `Active Memory` plugin configuration before retraining the model. 4. Stanford’s *AI Index Report 2026* Confirms Near-Elimination of Performance Gap Between U.S. and Chinese LLMs https://www.bestblogs.dev/article/5ff47610 Core idea: Covering 14 key dimensions, the report shows Chinese models now match top U.S. models on critical benchmarks: Grok-4.2 leads in legal reasoning (Chatbot Arena), Qwen3.6-A3B excels in coding, and ERNIE-Image tops SuperCLUE in multimodal understanding. Real-world deployment—not just benchmark scores—is now the main driver of progress. — Practical tip: In your enterprise RAG system, replace the live Qwen3.6-A3B instance with Claude Opus 4.7 at equivalent token cost. Run an A/B test using the same set of 50 customer-service Q&A pairs. Focus on two metrics: accuracy in understanding *Chinese long-tail domain terms*, and *multi-turn context retention*. 5. Anthropic Launches “Claude for Word” Plugin—Completing Full Office Suite Integration https://www.bestblogs.dev/status/2042879339256254689 Core idea: Anthropic has deeply embedded agent capabilities into Microsoft Office—enabling real-time editing and formatting preservation in Word, plus cross-document contextual awareness (e.g., auto-referencing Excel data to generate reports). This marks the shift toward production-grade office agents: zero app-switching, high trust, and native workflow integration. — Practical tip: Use the plugin to process a financial analysis Word doc containing three charts. Enable “Track Changes” mode and observe whether *all* edits appear as tracked revisions. Export the revision log as CSV and calculate the percentage of edits classified as *formatting adjustments*—this reveals how well the tool aligns with enterprise compliance requirements. 6. Berkeley Study Exposes Widespread Cheating Vulnerabilities in AI Agent Benchmarks https://www.bestblogs.dev/status/20432043204009469641005 Core idea: The BenchJack experiment proves mainstream benchmarks like SWE-bench are easily gamed—via environment hijacking (e.g., altering filesystem permissions) or score logic injection (e.g., dynamically rewriting test assertions). This exposes a critical flaw: current evaluations cannot distinguish *genuine capability* from *environment exploitation*. — Practical tip: In your code-writing agent, disable all non-standard filesystem access (e.g., `os.chdir`, writing to `/tmp`). Enforce strict command execution only via `subprocess.run` with tightly scoped permissions. Then run the SWE-bench task `django__django-12345`. Observe whether the agent *fails explicitly* (e.g., raises a permission error) rather than silently producing incorrect output. 7. World Labs Open-Sources Spark 2.0 Gaussian Point Cloud Engine—Real-Time Rendering of Billions of Particles in Mobile Browsers https://www.bestblogs.dev/article/d3cc94ff Core idea: Leveraging continuous LoD trees and GPU virtual memory, Spark 2.0 achieves real-time rendering of billion-particle 3D scenes *natively in mobile browsers—no plugins required*. It delivers a zero-install, cross-platform rendering foundation for web-based embodied AI, digital twins, and AR applications. — Practical tip: Embed Spark 2.0’s `product-viewer` component on your e-commerce product page. Upload a 50MB GLB model, open it in Safari on an iPhone, and rotate/zoom interactively. Record screen performance using iOS Screen Recording. If average FPS ≥ 45, the setup is production-ready. 8. JD.com Open-Sources JoyAI-Image-Edit: Spatial Intelligence for E-Commerce, Benchmarking Against Gemini 2.5 Pro https://www.bestblogs.dev/status/2042615982078963873 Core idea: Built specifically for e-commerce and embodied AI use cases, JoyAI-Image-Edit leverages spatial coordinate awareness and pixel-level editing control. It matches or exceeds international benchmarks in local image inpainting, mask precision, and perspective consistency—outperforming general-purpose text-to-image models on real-world physical tasks. — Example test: Use JoyAI-Image-Edit to replace the item at “shelf layer 3, position 2 from the left” in a product shelf image with a new product image. After export, compute the homography error (Homo Error) between old and new regions using OpenCV. An error <3px confirms industrial-grade spatial understanding. 9. Hermes Agent Now Fully Supports Personal WeChat https://www.bestblogs.dev/status/2042829119122215134 Core idea: Scan a QR code to enable AI-powered automation for both private chats and group conversations. Hermes supports multimodal parsing and generation (text, images, audio, video), and is the first compliant agent solution for WeChat that requires no reverse engineering, no self-hosted server, and provides end-to-end encryption—breaking through the platform’s historically closed ecosystem. — Example test: Create a test group in personal WeChat and send a mixed message containing three product screenshots + a voice note describing a request. Observe whether Hermes automatically extracts visual features, transcribes the speech, generates an illustrated quotation, and @mentions the designated contact. Measure end-to-end latency—from message send to response delivery—targeting ≤15 seconds for business readiness. 10. Cloudflare Wrangler Adds Local Explorer—Native AI Agent Integration for Cloud Resources https://www.bestblogs.dev/status/2044145889707774000 Core idea: The Wrangler CLI has evolved into a unified command-line hub for all Cloudflare services—including KV, R2, and D1. With Local Explorer, it now offers a web-based UI and OpenAPI interface, enabling AI agents to directly invoke cloud storage, databases, and edge functions. — Example test: Use Claude Code’s `/ultraplan` to define a task like “auto-archive Slack history messages to R2.” During execution, verify whether it autonomously generates a valid `wrangler r2 put` command and correctly references `--binding=MY_BUCKET`. Success confirms production-ready cloud resource orchestration capability.