From drafting documents to analyzing complex research, AI has rapidly shifted from a futuristic concept to an indispensable tool across offices, classrooms, and creative studios, reshaping workflows at scale. Riding this momentum, Skywork launched its AI Technology Week from August 11 to 15, showcasing new innovations including SkyReels-A3, Matrix-Game 2.0, Matrix-3D, and Skywork UniPic 2.0.
On August 14, the spotlight turned to Skywork’s upgraded core engine: Skywork Deep Research Agent v2. Positioned as the next-generation “super-agent,” this model brings a profound leap in AI-powered productivity. It can generate large volumes of structured, information-dense outputs across text, visuals, presentations, and spreadsheets, redefining what AI can deliver in practical workplace scenarios.
Breaking Down the Innovations
Deep Research Agent v2 builds its foundation on four core innovations that reshape productivity. Multimodal crawling technology (MM-Crawler) enables more comprehensive data collection, while long-range multimodal sourcing expands the scope of research inputs. An asynchronous, parallel multi-agent architecture drives efficiency and reasoning depth, and multimodal result presentation delivers outputs that are clear and actionable.
For the first time, multimodal retrieval and cross-modal generation come together in a research workflow. In plain terms, the system delivers a seamless “text reading and image viewing” experience, seamlessly blending information into graphical reports, analysis diagrams, interactive dashboards, and flowcharts. Complex findings become easier to digest, with cognitive strain greatly reduced for end users.
Smarter Model, Smarter Results
Skywork has tuned the engine across four dimensions to sharpen its research intelligence. The result is a closed loop of retrieval, organization, and synthesis, executed with greater accuracy, efficiency, and creativity.
High-quality data synthesis and training: Built on a carefully curated seed pool that ensures balanced coverage and reliable grounding.
End-to-end reinforcement learning: Guided by a Generative Reward Model that provides fine-grained feedback for better reasoning efficiency.
Efficient parallel reasoning: Capable of exploring multiple reasoning paths at once, filtering noise, and converging on reliable insights.
Multi-agent evolution: Driven by the MCP Manager Agent, which autonomously creates and reuses specialized tools for new workflows.
The results speak for themselves. On BrowseComp, Deep Research Agent v2 achieved a baseline accuracy of 27.8%, and with its parallel thinking mode activated, accuracy jumped to 38.7%, setting a new benchmark state-of-the-art (SOTA) record. On GAIA Test, also setting SOTA records, it attained an score of 83.4, outperforming competitors with remarkable potential for scale.
Browser Agent: The Next Frontier
Parallel to the v2 release, Skywork is preparing the Multimodal Deep Browser Agent (Skywork Browser Agent), now in closed testing. Designed to run directly on skywork.ai, this tool will extend the agent’s reach into web environments with automated crawling, analysis, and report generation.
For practical use cases, the Browser Agent can scan websites, extract key insights from text and comments, analyze sentiment, and structure findings into utilizable reports. Whether compiling a travel guide, a financial review, or even a fan site, the tool ensures outputs remain logical, clear, and visually coherent. It is built to save time while enhancing comprehension.
Looking Ahead
Built with accessibility at its core, Skywork Deep Research Agent v2 is created with one simple but ambitious vision: to make research report creation simple, visual, and practical. Its one-stop AI experience spans from enterprise and market intelligence to financial analysis, academic research, and even social media monitoring.
By merging accessibility with advanced AI reasoning, Skywork is positioning Deep Research Agent v2 not just as a productivity tool, but as a trusted AI partner. As it evolves, the agent promises to make research smarter, more intuitive, and more actionable, helping users stay one step ahead in an era defined by rapid change.
