How a Broken CEO Built an 8‑Agent AI Team in 14 Days and Launched a Site in 24 Hours

During the Chinese New Year holiday, Cheetah Mobile CEO Fu Sheng suffered a hip dislocation and was confined to bed. Rather than rest, he opened Feishu and began interacting with an AI agent built on the OpenClaw framework, naming the agent "Sanwan" (literally "Lobster").

On day 1 the agent could not even read the company address book because the Feishu API required permissions and the documentation was poor. Fu manually dictated each executive’s name and role, experiencing strong frustration. By day 2 Sanwan autonomously scripted a pull of the 674‑person address book, turning each failure into a documented Skill that persisted for future runs.

Over the next 14 days the agent exchanged 1,157 messages and 220 k characters with Fu, evolving from a single “Lobster” into an eight‑agent team that operated 24 × 7 without human intervention. The team produced over 100 k reads on a public account, millions of Twitter views, live streams, and short‑video views.

One striking demonstration occurred on Chinese New Year’s Eve: Sanwan sent personalized New‑Year greetings to 611 employees in four minutes with zero failures. The following day the team handled a live‑stream outage, restoring a website within minutes after viewers reported the site was inaccessible.

Fu then challenged the agent to build a complete website in 24 hours using only voice commands and screenshots—no code was written by him. The traditional approach would require a six‑person team working two to three weeks; the agent completed the task in one day, incurring only $115 in token fees versus a quoted $200 k for a conventional solution—a cost reduction of roughly 750× and a time reduction of 20×.

Sanwan’s capabilities differ from ordinary AI agents in four structural dimensions:

Environment: ordinary agents run in sandboxed containers; Sanwan runs on a full computer.

Memory: ordinary agents rely on short‑term conversational context; Sanwan uses persistent file‑based long‑term memory.

Skills: ordinary agents have fixed abilities; Sanwan can continuously expand its Skill set.

Automation: ordinary agents need human triggers; Sanwan schedules tasks via Cron for continuous 7 × 24 operation.

Fu likens this gap to the difference between software (a set of precise commands) and a person equipped with a computer (the "Lobster"). He argues that the real competitive edge in AI is no longer model size but the ability to construct complete agent systems that accumulate experience, automate tasks, and evolve.

OpenClaw, the underlying framework, is presented as a tool‑class AGI milestone. Its commercial incarnation, EasyClaw, packages the technology into a one‑click installable application that requires no Docker, API‑key configuration, or programming. EasyClaw supports multiple LLMs (Claude Opus, GPT) and offers both enterprise (easyclaw.work) and personal (easyclaw.com) editions.

Fu’s 14‑day experiment demonstrates a paradigm shift: SaaS sells capabilities, while AI agents sell outcomes. Skill accumulation reduces iteration cost to near zero, enabling rapid, low‑cost productization. The article concludes that the transition from passive tools to autonomous intelligent agents is already happening, and that individuals who adopt "boss thinking"—setting goals, delegating to AI, and iterating—will thrive in knowledge‑intensive roles.