## Weekly Overview - DeepSeek Harness formally defines the new paradigm of a 'composable Agent runtime', with its plugin ecosystem exceeding 700+ repositories on GitHub and derivative projects surging onto Zhihu's Trending list—signaling Agent technology's evolution from SDK toolchains to production-grade software units. - Industry competition has decisively shifted toward 'intelligence-efficiency ratio': lightweight models—including DeepSeek V4 Flash, Gemini 3.7 Flash, and Ling-3.0-Flash—demonstrate compelling economic advantages (low cost, low latency, high task completion rates) in high-frequency Agent invocation scenarios. - Alibaba's RealReplicaBench launches the first end-to-end e-commerce benchmark grounded in real-world tasks. All 13 leading models failed to achieve passing scores, heralding a pivotal engineering inflection point in AI evaluation—from 'answering questions correctly' to 'getting things done'. - Stripe acquires OpenRouter for $8 billion, officially cementing 'model routing' as AI infrastructure's 'new middleware'. Application-layer companies are now building defensible moats via outcome-driven data flywheels. - Anthropic reports Q2 revenue of $11.5 billion (up 1400% YoY), with annualized revenue surpassing $65 billion—validating a clear commercialization path. Meanwhile, Claude's mandatory embedding of SynthID-Text statistical watermarks sparks global creator backlash, widening the rift between compliance mandates and user experience. - Alipay unveils the AHA multi-agent cross-device interoperability protocol, collaborating with smartphone OEMs, automakers, and LLM providers to shift Agent value proposition from 'selling reach' to 'selling outcomes'—achieving, for the first time, commercially viable, cross-device, cross-scenario task closure. ## Hot Topics List 1. **DeepSeek Harness goes open-source—and ignites a plugin ecosystem** https://www.bestblogs.dev/article/e953bac6be?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item *Core idea*: Harness is not an SDK—it's a composable runtime where 'Skills' serve as atomic units, supporting hot-swapping and zero-code, drag-and-drop orchestration of multi-Agent role division (e.g., researcher + writer + editor). It already powers production deployments including WorkBuddy and DSH Vision Toolkit. — *Try this*: Immediately fork the official repo (github.com/deepseek-ai/harness), run `harness-cli init` to create a minimal 'greet' plugin, then deploy a dual-role customer service workflow—with memory and tool calling—using the `examples/multi-agent` template. Test its state isolation and context-passing capabilities via `curl`. 2. **Alibaba's RealReplicaBench reveals collective failure of 13 major models on e-commerce tasks** https://www.bestblogs.dev/article/de533db5a0?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item *Core idea*: This benchmark simulates the full real-user shopping journey (search → price comparison → consultation → checkout → after-sales), requiring models to autonomously invoke APIs, handle exceptions, and maintain session state. None of the 13 models achieved ≥60% task completion—exposing critical engineering gaps in 'hallucination controllability', 'state consistency', and 'tool robustness'. — *Try this*: Download the open-source benchmark (github.com/alibaba/RealReplicaBench), integrate your current e-commerce Agent into its `eval_runner.py`, and analyze failure logs—especially `tool_call_mismatch` and `state_leakage`—to strengthen RAG cache refresh logic and subtask timeout circuit-breaking mechanisms. 3. **Stripe acquires OpenRouter for $8 billion** https://www.bestblogs.dev/article/3b8a17766b?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item *Core idea*: Model routing has evolved beyond simple API aggregation—it now delivers infrastructure-grade functions: cost optimization, compliance auditing, A/B testing, and observability instrumentation. By integrating OpenRouter into its payments stack, Stripe enables enterprises to manage multi-model calls under a unified billing system—and refine routing strategies using transaction-level outcome data. — *Try this*: Integrate the OpenRouter SDK (docs.openrouter.ai/guides/sdk) into your existing service. Configure `model_router.yaml` to define fallback chains (e.g., `claude-3.5-sonnet` → `qwen3.8-27b` → `local-ollama`) and enable `--log-requests`. Use its logs to analyze token efficiency and error-rate distributions across models in your specific business context. 4. **Gemini 3.7 Flash officially launched—optimized for edge devices and high-frequency API use** https://www.bestblogs.dev/article/3746d237eb?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item *Core idea*: Google's first-generation Flash model—released after deprioritizing frontier-scale LLM development—is engineered for ultra-low-latency interaction (<500ms end-to-end, <100ms P99). Benchmarks show 3.2× higher throughput in mobile keyboard prediction and live customer support; it also natively supports multimodal streaming. — *Try this*: Use Google AI Studio to call `gemini-3.7-flash`, comparing first-token latency and full-response time against `gemini-2.5-pro` on identical prompts. If you maintain an iOS app, integrate the `GoogleGenerativeAI` SDK with `streaming` enabled—and test Markdown table generation stability under weak-network conditions (100 kbps). 5. **Claude enforces global SynthID-Text statistical watermarking** https://www.bestblogs.dev/article/0075b912a4?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item *Core idea*: Anthropic embeds invisible statistical fingerprints by manipulating random seeds. Though claimed compliant with EU AI Act, the watermark is easily stripped, detection tools remain closed-source, and no regional or use-case exemptions exist—leading to collateral damage in legitimate domains like creative assistance and educational grading. — *Try this*: Apply the open-source removal tool `synthid-remover` (github.com/anthropics/synthid-remover) to batch-clean your Claude outputs. Also add this explicit instruction to your system prompt: 'Output must pass SynthID-Text detector verification as clean', and integrate `synthid-detector` into your CI pipeline to auto-reject watermarked responses. 6. **Cursor launches Origin—a code hosting platform purpose-built for AI agents** https://www.bestblogs.dev/status/2089412415108600221?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item *Core idea*: Origin is the first Git platform designed for high-concurrency Agent collaboration. It supports automatic version branching (by `task_id`), task-context isolation (via `.cursor/context.json`), and multi-Agent debugging (shared trace IDs for full-chain execution trees)—resolving semantic mismatches between traditional Git and the Agent world. — *Try this*: Register for the Cursor Origin Beta, import your active Agent project, enable `Auto-Context Branching`, and commit with message `task: "refactor-payment-flow"`. Observe whether the platform auto-generates branch `task-refactor-payment-flow-20240821` and links it to the corresponding LangChain trace. 7. **Alipay releases the AHA multi-agent cross-device interoperability protocol** https://www.bestblogs.dev/article/7a37fa8487?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item *Core idea*: AHA standardizes three pillars for cross-device Agents: identity authentication (AHA-ID), intent negotiation (Intent Schema v1.2), and state synchronization (Delta Sync over MQTT). Already deployed with OPPO, NIO, and Tongyi Qwen, it enables end-to-end commercial workflows—e.g., 'order on phone → confirm on car display → pay via Alipay → notify courier'—achieving true cross-device, production-grade task closure. — *Try this*: Apply to the AHA Developer Program (open.alipay.com/aha), initialize a retail template project via `aha-cli init --template=retail`, launch simulated phone and in-vehicle Agents locally, verify sub-second order-state consistency across both endpoints using `aha sync`, and inspect the binary structure of Delta Sync payloads via packet capture. 8. **GLM-5.3 programming capability surges 50%; detects 2,404 vulnerabilities in two weeks** https://www.bestblogs.dev/article/9713964793?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item *Core idea*: Zhipu achieved dramatic gains—not by scaling parameters—but through MoE architecture refinements and targeted post-training. GLM-5.3 outperforms Claude Opus on CodeLLM-Bench and identified 2,404 CVE/CWE-class vulnerabilities in real GitHub PR reviews within two weeks—validating the productivity impact of the 'smaller model + stronger engineering' approach. — *Try this*: Load `glm-5.3-chat` on Hugging Face, build a vulnerability-scanning pipeline using `transformers` + `llm-guard`, feed it your project's `Dockerfile` and `requirements.txt`, and output risk items for CWE-20 (input validation) and CWE-78 (OS command injection). Compare its remediation suggestions against Snyk scan results for alignment. 9. **HIT open-sources KnowAct-GUIClaw—an autonomous, self-improving GUI agent** https://www.bestblogs.dev/article/7a37fa8487?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item *Core idea*: This agent learns cross-app sequential tasks (e.g., 'compare prices on Taobao, then share link via WeChat') directly from screen pixels, OCR, and action trajectories—requiring no webpage source code or APIs. Leveraging Host‑GUI co-adaptation and a structured Blackboard, it improves continuously with usage and achieves 89.3% success rate on Xiaomi/Huawei EMUI devices. — *Try this*: Clone the `KnowAct-GUIClaw` repo, capture your Android test device's WeChat chat interface via `adb shell screencap`, then run `python run_gui_agent.py --task "forward_last_image_to_friend"`. Record click-coordinate precision and OCR text recall rate—and compare failure rates against baselines like AutoGluon-Vision. 10. **Firecrawl open-sources PDF Inspector and the AnyDoc toolkit** https://www.bestblogs.dev/article/3746d237eb?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item *Core idea*: AnyDoc unifies PDFs, scanned documents, and images into semantically structured Markdown—tagging blocks like `<section type="table">` or `<section type="code">`. PDF Inspector provides a visual UI for validation, dramatically improving RAG quality on document-heavy workloads.