Intel first proposes intelligent PC concept as digital employee era sees “Lobster” and “Hermes” surge

On April 22, Intel China’s technology director Gao Yu said the “Lobster” product exploded in popularity from March, and although enthusiasm cooled in April, product evolution has not slowed; he described the technological revolution driven by intelligent agents as irreversible. After the phenomenon of “Lobster” OpenClaw, another open‑source agent, “Hermes” (Hermes Agent), has recently gained traction in the tech community. Industry sources told Cailian Press that the rapid rise of “Lobster” and “Hermes” has accelerated the adoption of digital employees, with 2026 being called the first year of digital employees; these digital employees can autonomously perform duties 24/7 and further promote the One Person Company (OPC) model.

On the software side, vendors are both adapting popular open‑source agents and launching proprietary agent solutions. Last week, Tencent Cloud’s Lighthouse became the first in the industry to enable one‑click cloud deployment of “Hermes,” while MiniMax launched MaxHermes, a cloud‑based, self‑evolving AI assistant built on Hermes Agent that claims users can have an agent that “learns them better over time” within ten seconds. At the same time, Lanzhou Technology released a trusted AI full‑stack system covering an enterprise agent OS (LangClaw), digital employees, an enterprise brain and foundational trusted capabilities.

Hardware makers are also seizing the opportunity, with multiple “Lobster” boxes and all‑in‑one machines emphasizing secure local operation and addressing users’ token cost anxieties. On April 21, Intel — which first proposed the AI PC concept in September 2023 — formally introduced the “intelligent PC” concept to address security risks, high costs and fragility exposed by agents.

The surge of “Lobster” is accelerating the emergence of “super individuals.” Lanzhou Technology founder Dr. Zhou Ming told Cailian Press that China’s AI industry has entered a golden development period, recalling that 2023 was the year of conversational AI, 2024 the year of reasoning AI, 2025 the year of agent AI, and 2026 the year of digital employees. On April 21, the State Council issued guidance to advance “AI+” initiatives, accelerate intelligent programming tools and support procurement of large models and agent services; a prior policy set a target of over 70% application penetration for agents and related applications by 2027.

Agents are moving rapidly into large‑scale deployment. Zhou described the technical path as evolving from large‑model enhancement to mature agent engineering and then to digital employee implementation. Agent engineering uses external data with fine‑tuning and RAG to strengthen agents, and employs skills, MCP and A2A protocols to call on tasks or other agents for complex workflows; prompt engineering, context engineering and memory optimization further improve task understanding, decision support and memory. From this methodology, an AI‑native software architecture is emerging: unlike traditional AI+ approaches that merely add capabilities and remain loosely integrated with business systems, AI‑native designs around LLMs and agents to enable autonomous decision‑making, process automation and deep business collaboration.

Zhou said AI‑native concepts and agent engineering form the technical foundation for the OPC model, which has already received policy support in more than 20 cities. However, Cailian Press noted that some early OPC startups have failed, underscoring that lower entry barriers do not guarantee success; many products lack core differentiation and face monetization challenges, with reports of nearly 30% vacancy in some OPC coworking communities. Zhou acknowledged that startup failure is common and said Lanzhou Technology aims to provide digital employee skills for sales, marketing and recruitment to help OPC entrepreneurs address customer acquisition and hiring pain points.

LangClaw is positioned as a unified runtime for digital employees, supporting natural‑language task conversion, cross‑system operations, multi‑employee collaboration, secure sandbox execution and full‑chain auditing. Lanzhou Technology said it will continue to deepen AI‑native and trusted AI capabilities, focusing on finance, telecom and state‑owned enterprises, and use LangClaw to scale agents from pilot projects to full business coverage.

On April 21, Intel formally proposed the intelligent PC for personal scenarios, complementing enterprise inference systems and containerized agent farms to form a full‑stack agent platform. The intelligent PC is optimized for agent use, offering local task closure, hybrid edge‑cloud inference, memory and self‑evolution, and local security protections. Product forms include laptops, mini PCs, all‑in‑ones, AI boxes, AI NAS and edge gateways, using a “local assistant brain + cloud main brain” hybrid architecture that keeps sensitive tasks local while invoking cloud resources for high‑capability tasks. Intel outlined flagship configurations with third‑generation Core Ultra X processors and 32GB+ memory capable of running large models such as Qwen3.5/3.6 (35B) and Gemma4 (26B), mainstream configurations with third‑generation Core Ultra 8‑ or 16‑core processors and 16GB+ memory for mid‑sized models like Qwen3.5 (9B/4B), and entry configurations with third‑generation Core processors and lower memory (e.g., 12GB) for lightweight models such as Qwen3.5 (4B) to handle multimodal tasks like PaddleOCR and CosyVoice. Gao noted that AI SSDs further improve performance for memory‑constrained systems, citing Phison’s aiDAPTIV technology as enabling thin laptops with 16GB memory to run Qwen‑35B‑class models, and he said the March “Lobster” boom tightened supply of x86 PCs, which he described as the true AI‑native architecture.

To enable out‑of‑the‑box agent experiences, Intel is building an AI Skill ecosystem with partners including Cherry Studio, Yingtao AI Assistant, DuMate, Flowy, Juxin Technology, Molili, QClaw, QoderWork, QwenPAW, Jieyue Xingchen, Tencent App Store, TRAE and Workbuddy. Intel has developed selected Skills for high‑frequency scenarios such as intelligent search, quick video editing, game assistance and security guardrails, and has open‑sourced reference Skills for developers while encouraging community contributions through the Moda community AI PC zone and innovation competitions.

The agent ecosystem’s rapid growth has driven exponential increases in token consumption, raising cost concerns. National Data Bureau director Liu Liehong disclosed at the China Development Forum 2026 that China’s daily token calls rose from 100 billion in early 2024 to 100 trillion by the end of 2025, and exceeded 140 trillion in March 2026, a more than thousandfold increase over two years. Analysts attribute the surge to large‑scale image and video generation by products such as Seedance and Nano Banana, and to multi‑step agent workflows like those executed by “Lobster,” which run at high concurrency and require continuous iteration.

Zhou explained that while single‑round AIGC queries already consumed significant tokens, agents have multiplied token usage by enabling many parallel agents that work continuously and autonomously gather and summarize information, making token consumption sometimes difficult to control. He said token‑saving strategies are a hot topic and a necessary consideration when building agent OSs, and that hybrid orchestration of large and small models across cloud and edge can reduce token costs. Intel’s intelligent PC, with its hybrid deployment model, is also designed to address token cost concerns by balancing expensive cloud calls with local processing.

From a total cost‑of‑ownership perspective, Gao said the optimal model size for edge deployment is around 35B parameters, as larger models raise costs and smaller models face capability limits. A recent A‑share technology company released an all‑in‑one machine that supports “Lobster” and, by enabling local deployment of models and Claw applications, can reduce token anxiety and lower long‑term costs by up to 60% compared with continuous cloud usage. Intel acknowledged that “Lobster” is not aimed at ordinary consumers but has promoted agent adoption. Gao concluded that reducing token consumption and lowering usage barriers are necessary conditions for mass acceptance, while other challenges will require time to resolve.