Tencent's 78-Page Agent Handbook: Chapter 1 Puts the Three Things Enterprises Fear Most on the Table

Tencent, Tencent Cloud, and WorkBuddy jointly released a 78-page '2026 AI Agent Enterprise Solutions and Implementation Practices Handbook.' meng shao said on X that he plans to start reading it chapter by chapter over the weekend; Chapter 1 is 'AI Overall Solutions.' So far, only a paid download address has been found for this PDF, with no free public link. In his tweet he promised to send it by DM to anyone who needed it, and the comments were soon flooded with 'please send me a copy.'

Chapter 1 does this: it folds Tencent Cloud's sprawling set of Agent products into one chain. The entry point is WorkBuddy, positioned as an office workbench and digital employee; below that, ClawPro handles provisioning and unified control; Agent Runtime provides the sandbox, gateway, and governance foundation; at the very bottom, TokenHub connects models in a unified way. The value proposition in one sentence: organizational capacity shifts from 'human collaboration' to 'humans set goals plus digital employees deliver.'

The panorama is even denser. On the personal side there are LightClaw and QClaw; on the office side WorkBuddy; on the R&D side CodeBuddy; for enterprise applications ADP×ClawPro; plus Lexiang, Qidian Marketing Cloud, and Data Agent. The foundation is TokenHub (Hunyuan plus third-party models) and Agent Runtime (execution engine, cloud sandbox, memory, gateway, observability), with security center, AI Agent security gateway, and iOA attached on the periphery. SkillHub claims 25,000+ ecosystem open-source Skills. The official main line is called 'from super individual to super team.'

The WorkBuddy page presents a three-stage argument: can converse, can do work, can take a post. Q&A tools count as the first stage; an office assistant embedded in workflows and able to call tools is the second stage; a digital employee must enter business systems, change data, and be responsible for results and KPIs to count as the third stage. The accompanying 'three generations of office work' formulation is: humans operate software; humans chase and fill in forms inside collaboration platforms; humans set goals, set boundaries, and make judgments. The handbook says that the more complex the system and the denser the processes, the more it becomes the battlefield for digital employees. Read in reverse, it also holds: in places where processes are loose and systems are shallow, the returns from deploying Agents first are limited.

ClawPro is the control console for enterprise IT. Employees can activate with one click; both open-source and self-developed Agents can be connected; security has four layers—see, audit, manage, and scan; it supports multiple IM access points including WeChat, QQ, WeCom, Feishu, and DingTalk; it supports grouping and Token/Agent quotas; Claw Team handles multi-Agent orchestration; and there is a private Skill library. In the comments, a netizen working on a baijiu cloud said that what often blocks enterprises from buying Agents is not capability, but IT's three questions: who created it, which data it can touch, and who gets investigated when something goes wrong.

Chapter 1 lays out the functions corresponding to these three questions, but it does not fully explain how responsibility is divided. Another netizen, Yunqi, asked more directly: if digital employees are responsible for KPIs, who does the error count against—does the handbook go that far?

The Agent Runtime section is where technical readers should pause. It treats the Agent as the primary perspective, and beyond the governance platform layer there are seven atomic capabilities: Sandbox, Wall, Gateway, Identity, Observability, Skills, and Memory. Sandbox isolation, outbound/inbound control, zero trust plus credential hosting and rotation, model and tool gateway auditing, and full-link observability. A netizen asked whether this layer is self-developed by Tencent Cloud or can be compatible with other vendors' frameworks; the material does not give an answer.

The most practical thing in TokenHub for developers: the same Base URL plus the same API Key is compatible with both OpenAI and Anthropic protocols. Clients such as Claude Code, Cursor, Cline, and OpenClaw only need to change one line of configuration to switch models. Some netizens also reminded that switching models is easy; what really blocks enterprises is permissions and auditing at the Runtime layer, and digital employees that can change business data are most afraid of a situation where something goes wrong and cannot be rolled back.

There is another detail. SkillHub says it has 25,000+ open-source Skills; the number sounds nice, but if enterprises really want to use them, someone has to first review what is written inside each skill. That workload is not written into the handbook.

The case numbers in the handbook—such as content efficiency close to 80%, MTTR of 5 to 15 minutes, operations and maintenance costs down 40%, and repetitive O&M reduced by 70%—are all vendor self-reported and have not been third-party verified. Some netizens say it is more accurate to read it as a product catalog. Others asked why the official original source cannot be found; meng shao's reply was that he only found a paid download link. The common problem with this kind of handbook: the chain is drawn very completely, but the boundaries and failure paths are left for readers to fill in themselves.

Different readers can pick and choose. Those with a development background should read the Agent Runtime and ClawPro sections first; permissions, auditing, and rollback determine whether it can really go into production. Those without a technical background will find the WorkBuddy section's path from 'can converse to can do work' most practical: first find a small scenario embedded in a workflow and get it running end to end, and do not start by wanting a digital employee. Those doing procurement should copy IT's three questions into a checklist and ask suppliers one by one.