Source | Benyuan Finance

Author | Li Youshan

In 2026, the most dynamic arena for AI is undoubtedly in the office sector. As models evolve to become increasingly versatile, the competitive focus among tech giants has shifted to a deeper level, transforming into an all-out "dual-front" war encompassing both software and hardware: personal agents have surged in popularity, while AI hardware continues to emerge in abundance.

From Anthropic's launch of Claude Cowork at the beginning of the year, setting the pace for personal agents, to Meta, Alibaba, and ByteDance's dense (Note: ' dense ' is translated as 'frequent' or 'intensive' in context, but kept as pinyin here due to the request for HTML format retention; in a real scenario, consider replacing with appropriate English term) deployment of intelligent hardware by the end of the year, leading companies are using software agents to reshape digital office workflows while simultaneously leveraging hardware terminals to capture entry points for perception in the physical world. The speed of product iteration even outpaces the update rhythm of workplace tasks.

As AI office solutions officially transition from "conversational tools" to "task execution," the core industry divergence (Note: ' divergence ' is translated as 'divergence' in context, but kept as pinyin; replace with appropriate term as needed) has become increasingly clear: which track—software-based personal agents or diversified AI hardware—holds greater long-term potential?

Making AI More Human-Like

Over the past year, AI agents have collectively stepped out of chatboxes and plunged into real-world business scenarios, with their numbers rapidly expanding. According to IDC, the number of active agents is expected to skyrocket from approximately 28.6 million in 2025 to 2.216 billion by 2030.

In the AI office domain, user interest was ignited by last year's Manus and this year's OpenClaw market education efforts. The focus of competition among tech giants has shifted from vying for entry points to actual task execution, culminating in the recent emergence of "personal agents" capable of handling long-term affair (Note: ' affair ' is translated as 'tasks' or 'affairs' in context; kept as pinyin here).

Anthropic, an early adopter of office agents, launched Claude Cowork for knowledge workers in January 2026, enabling direct access to local files, coordinating sub-agents for parallel task processing, and autonomously executing complete workflows. In June, OpenAI integrated Codex into ChatGPT while introducing Agent Plugins and Sites, bundling agent capabilities for code, documentation, and data processing into users' existing habits. September marked a concentrated outbreak (Note: ' outbreak ' is translated as 'breakout' or 'surge' in context; kept as pinyin) period for the global personal agent sector.

Meta fired the first shot with Muse, which quickly became the year's phenomenal product. In just 13 days, Meta topped the U.S. Apple App Store's free applications chart, driving an 11.43% surge in Meta's stock price on the same day. Muse broke away from the traditional question-and-answer interaction paradigm of conversational AI, continuously remembering user information and understanding long-term intentions.

During the product launch, Muse demonstrated a daily scenario: when a user bookmarked a recipe on Instagram, Muse would compile a menu and shopping list based on family members' dietary restrictions, then proceed with related tasks. Its popularity in North America validated the mass-market acceptance of personal agent products among consumer audiences.

Manus followed suit with the release of its personal agent application, Cue. Cue equipped each agent with an independent email address, phone number, wallet, and computing terminal, granting it a complete digital identity. It could autonomously send and receive messages, make payments within budget, independently execute tasks from start to finish, answer calls on behalf of users, and automatically generate call summaries, further expanding agents' responsibilities and application scenarios.

At OpenAI's "largest-ever" developer conference in late September, Sam Altman dedicated the finale to Dots. OpenAI described Dots as "agents capable of continuously assuming responsibilities," emphasizing an "always-on" feature and proposing a new direction for future teams of multiple dots to collaborate.

The primary market's enthusiasm soared in tandem. That same month, Instinct, a startup with only 14 employees, secured $1 billion in funding, propelling its valuation past $10 billion. Instinct employs SMS and phone interactions, currently offering product access by invitation only. Overseas companies collectively view personal agents as the next-generation traffic entry point following search engines and app stores, while domestic tech giants have also fully rolled out their strategies.

In late September, ByteDance was reported to be accelerating its personal agent project, codenamed Spell, led by the Doubao mobile assistant team. Tencent Yuanbao already offers an AI-native desktop intelligent agent workstation, WorkBuddy, along with the "cloud-based personal assistant" LightVela and the WeChat AI assistant "Xiaowei" (currently in testing), with recent rumors suggesting an imminent launch of a personal agent product.

Entering October, Ant Group's Lingguang initiated a new version of its internal test, focusing on enhancing personal task execution capabilities. Alibaba's Qianwen had previously announced accelerated development of its Personal Agent, officially launching the personal assistant "Qianwen Go!" on the 13th and conducting a limited-time invitation test on the Qianwen App.

Among startups, Today AI, founded by former Doubao executive Qi Junyuan, opened for downloads on September 24, emphasizing long-term memory and background parallel execution, earning it the nickname "China's Muse." With global tech giants and startups fully entering the fray, AI agents have officially transitioned from "conversational tools" to "task execution," shifting focus from "tasks" to "people."

Integrating AI Hardware into Daily Life

The evolution of large models and agents has prompted even top scientists to issue warnings, yet in the real world, most personal terminals remain unchanged. While personal agents address task execution in the digital realm, AI hardware serves as a crucial gateway bridging the physical and digital worlds.

A plethora of new AI hardware forms have emerged, from AI glasses and earphones to recording cards, necklaces, and rings, constantly pushing the boundaries of intelligent cognition under the rallying cry of "reinventing all hardware with AI."

The core function of hardware products is to act as AI's "ears" and "eyes" for understanding the real world, providing agents with more complete user contexts and addressing their "perceptual shortcomings," thereby truly integrating technology into daily human life.

Overseas tech giants pursue a dual strategy of advancing underlying computing power and terminal forms for AI hardware.

Intel unveiled the Core Ultra Series 3 processor at CES 2026, partnering with Microsoft to expand Copilot+ PC to a broader consumer market. Microsoft collaborated with NVIDIA to launch Windows laptops powered by the RTX Spark chip, capable of running high-performance AI models locally. Amazon introduced its first voice assistant-named Alexa series tablet, embedding AI assistants into hardware systems. Apple released HomeView, a smart home hub focused on end-side AI experiences for home and light office scenarios.

Domestic tech giants, meanwhile, prioritize hardware deployment in office scenarios to collect offline office data and address software agents' contextual shortcomings.

Take the fiercely competitive lightweight recording sector as an example. Plaud reignited the recording device war with its card-style AI summary product, Plaud Note. In September, Plaud announced a deep collaboration with Tencent WorkBuddy, interconnecting their official MCP interfaces to seamlessly integrate into workflows, transforming single conversations from "a summary" into reusable work assets. A co-branded hardware product, Plaud Note Pro, is also planned.

Tech giants swiftly followed suit. Alibaba DingTalk released its first AI recording hardware, A1, in 2025. This September, Qianwen Office launched the upgraded QwenNote A2, emphasizing a "burn-after-transcription" privacy feature. iFLYTEK, a veteran in the voice sector, entered the market in September with its AI recording card, leveraging its native voice technology advantages. ByteDance's Feishu partnered with Anker Innovations to launch the "AI Recording Bean" in early 2026, weighing just 10 grams and adaptable to daily interviews, meetings, and other scenarios.

AI glasses have also emerged as a standout category in AI hardware this year, becoming a focal point in the tech giants' arms race. Omdia data shows global AI glasses shipments surging from 380,000 units in 2023 and 2.06 million units in 2024 to 8.7 million units in 2025, with an estimated breakthrough of 15 million units in 2026, making it one of the fastest-growing consumer electronics categories.

In the vertical sector, manufacturers like Rokid, Xreal, and Thunderbird Innovation have taken turns claiming top shipment volumes, while collective entry by internet giants is propelling this track from geek circles to the mainstream market.

Alibaba stands out as the most aggressive domestic player. In March 2026, Qianwen officially launched its first AI glasses, G1, positioned as a "portable intelligent super assistant" fully integrated with Qianwen's large model and office ecosystem. In October, it released the next-generation Qianwen AI glasses, N1 and N1 Pro, with the N1 Pro supporting vital sign monitoring like heart rate and blood oxygen, further expanding perceptual dimensions. Shu Junliang, VP of Qianwen Office, stated, "Hardware serves as an entry point for incremental context collection, bridging the digital and physical worlds."

Xiaomi and Baidu have also joined the fray, with Xiaomi's AI glasses focusing on first-person filming and life logging, and Baidu emphasizing translation and portable AI assistants, both without screen displays. Tencent, through its WorkBuddy open platform, collaborated with Rokid to launch the Leqi Rokid×WorkBuddy co-branded AI glasses.

This represents just the tip of the hardware iceberg. To dominate the AI office sector, mine high-value data from offline scenarios, and complement physical world data inaccessible to pure large models, tech giants have extended their hardware strategies to encompass all forms of audio-visual wearables, igniting a full-scale battle for entry points.

Short-Term Software Surge, Long-Term Hardware Dominance

Personal agents and AI hardware represent two stages of AI office evolution, but significant differences exist in their growth potential, barrier height, and payoff timelines when viewed as independent tracks.

The short-term potential of personal agents is undeniable, with the steepest growth trajectory.

Software's near-zero marginal costs and high distribution efficiency ensure personal agents scale far faster than hardware. From a user experience perspective, agent capabilities directly tie to underlying large models, with product evolution measured in weeks or even days.

Commercially, personal agents inherently possess a payment foundation. Instinct founder Noah Shinn revealed its annualized transaction volume has surpassed $1 billion, translating to $30 million to $50 million in annual revenue at a 3% to 5% commission rate.

Personal agents are poised to become genuine "work-life companions" for daily use. However, their development bottlenecks are clear, facing challenges in commercialization, social compliance, and trust.

First, greater capability demands stronger judgment. Balancing answering, thinking, and acting determines whether personal agents save time or create new work.

Moreover, mainstream domestic apps bundle social, payment, shopping, and travel functions into super apps, offering convenient operations within walled gardens. WeChat, Taobao, Douyin, and Meituan restrict data sharing and permission access, directly capping agents' capability ceilings.

This means personal agents cannot achieve true cross-platform, full-scenario execution, remaining confined to "auxiliary tool" status in terms of authority boundaries and autonomous execution depth, far from "independent task handling."

As for privacy concerns, Muse faced a setback in September when nearly 190,000 user conversation records were leaked, forcing emergency data permission tightening and full-process authorization prompts.

Compared to personal agents, AI hardware collects exclusive offline data such as voice, scenarios, and vital signs, which software agents can never access, offering higher long-term value barriers.

Yet, it faces high R&D, supply chain, and channel costs, long iteration cycles, high "dust-gathering" rates for products, severe homogenization within single categories, and limited ceilings. It must overcome user habit cultivation and supply chain maturity cycles to prove sufficient rigid value.

Models address "how contexts are understood," agents tackle "how contexts are executed," and hardware solves "where contexts originate." Together, they coalesce into an integrated smart ecosystem.

The AI-driven industrial wave has long resonated across industries, leaving no one untouched.

From question-and-answer conversational tools to autonomous execution entities, from purely digital online scenarios to full coverage of virtual-real integrated office environments, this transformation is not only reshaping the underlying logic of productivity but also redefining the collaborative boundaries between humans and intelligence.

Operations/Yu Shuya Design/Yan Weier

*All rights reserved. No reproduction without authorization.

end