Tencent Handbook Breaks Down the Agent Technology Foundation: StepFun Post-Training Efficiency Up 30%
Chapter 3 of the 78-page "2026 AI Agent Enterprise Solutions and Implementation Practices Handbook," jointly released by Tencent, Tencent Cloud, and WorkBuddy, breaks the Agent technology foundation into a training-side elastic sandbox and an application-side Agent Runtime. The handbook claims the sandbox can be delivered in about 60ms, with a goal of spinning up about 160,000 instances of different images within 1 minute; StepFun used this to achieve roughly a 30% improvement in post-training efficiency, while Kimi Agent Swarm supports up to about 300 sub-agents and over 4,000 tool calls, with sandbox startup reduced to the order of about 80ms.
View original postTencent, Tencent Cloud, and WorkBuddy jointly released a 78-page "2026 AI Agent Enterprise Solutions and Implementation Practices Handbook." The entire document is tied together by a single product narrative: first, WorkBuddy pushes the personal assistant into a "digital employee" embedded in business systems; then ClawPro, Agent Runtime, and TokenHub handle provisioning, isolation, models, and governance; the remaining chapters cover how this foundation rolls out across industries including R&D and operations, SaaS, content, social, gaming, and customer service. Overall framework: three stages × three generations of office work · From conversation → to doing work → to taking a job: Q&A tools (the document cites ChatGPT / DeepSeek as examples) → office assistants embedded in workflows and calling tools (WorkBuddy) → digital employees that enter business systems, are accountable for results, and deliver against KPIs (Agent workbench + business system integration). · Three generations of office work: people operating software (Office) → people chasing and filling in forms within collaboration platforms → people setting goals, defining boundaries, and making judgments, while digital employees understand context, call tools, execute continuously, and deliver completed work. · "Super individual → super team": on the individual side, LightClaw / QClaw and others; on the enterprise side, WorkBuddy as the entry point, consolidating people, Skills, connectors, and the control plane into replicable execution capability. The document also distinguishes office work in the narrow sense (a personal productivity entry point) from office work in the broad sense (actually changing business state in CRM / ERP / MES). The handbook also provides detailed explanations and distinctions of the roles of WorkBuddy, CodeBuddy, ClawPro, ADP, SkillHub, and others within enterprises. Internal enterprise efficiency (Section 2 of the solutions book) · WorkBuddy: uses a "business brain" to assemble holistic judgments; for example, it operates real CRM/ERP/MES through Digiwin MCP, emphasizing that integration testing verifies changing state, not just being able to answer. There are many "before → after adoption" comparison pages for onboarding configuration, finance/sales/delivery dates, and more. On the security side, it claims to be among the first batch recognized by the Ministry of Industry and Information Technology, among the first batch for the CAICT cloud-based agent baseline, and "two-star+" for personal information protection impact assessment, among others. · CodeBuddy: from AI-assisted to AI-native; making specifications (SPEC) explicit; the requirements—coding—QA—delivery chain is underpinned by Agents, with human review and acceptance. · Game R&D / operations: planning and prototyping use WorkBuddy, R&D integration uses CodeBuddy; on the operations side, data Agents such as DataBuddy. The document claims planning cycles and demo cycles are significantly shortened (specific multiples are self-reported). · SRE / databases: WorkBuddy × CloudQ for fault localization; DatabaseClaw claims to have accumulated experience from hundreds of thousands of tickets, 1,600+ OpenAPIs, more than ten database types, and tiered handling. · Case names: Shenghe, Taomee (content efficiency near 80%, covering about 80% of employees), Youzu (47 models, Agent penetration over 75%, along with changes in overseas revenue/gross margin), Youka (MTTR 5–15 minutes, operations costs about -40%), Sohu Changyou (shortened P0 recovery, repetitive operations -70%, etc.), Huajiao, Zilong, Mabang, and others. Several chapters on business innovation · SaaS Agent: subscription → Credits → pay-per-result; interface operations move toward Skill/API. Names include MiningLamp, Entropy Simplify, Huisuanzhang, Futong JoinfClaw, BlueFocus, Six Degrees Renhe, Guangmo, Tiancai Shanglong, and others. · Content production: marketing images and videos, game art (Miora / Ardot, etc.); efficiency comparisons such as 2D sprite frames are self-reported in the document. Kuaizi, LiblibAI, Bifan, Wanxinbuzhi, Zixun, Xiaoying, Dongqiudi, and others. · Social companionship / in-game AI / customer service: TokenHub + audio/video and voice pipelines; NPC memory and tasks; customer service/quality inspection/outbound calling (Tianrun Rongtong, Qingxi Daijia, Zhichi, etc.). Latency and resolution rate figures are all as claimed in the document. · Ecosystem co-building: Buddy apps, experts, Skills, connectors, hardware; it claims to be a leader in desktop office agents, with tens of millions of installs, hundreds of thousands of ecosystem assets, and over a hundred partners (as self-reported in the document); PatSnap, Midu, MasterGo, and others appear on the partner pages. For the complete report, you can DM me to get it (not deliberately farming for engagement — it's because I couldn't find a free, publicly accessible PDF address; it's a paid download link, so DM me or leave a comment and I can send you the PDF).
View original
腾讯、腾讯云和 WorkBuddy 联合发布了一份 78 页的「2026 AI Agent 企业解决方案及落地实践手册」 整份材料用同一条产品叙事串起来：先用 WorkBuddy 把个人助手推成进业务系统的「数字员工」，再用 ClawPro、Agent Runtime、TokenHub 做开通、隔离、模型与治理；其余章节是这条底座在研发运维、SaaS、内容、社交、游戏、客服上的行业铺开。 总框架：三阶段 × 三代办公 · 能对话 → 能干活 → 能上岗：问答工具（文中举例 ChatGPT / DeepSeek）→ 嵌工作流、调工具的办公助手（WorkBuddy）→ 进入业务系统、对结果负责、按 KPI 交付的数字员工（Agent 工作台 + 业务系统集成）。 · 三代办公：人操作软件（Office）→ 人在协同平台里催办填报 → 人给目标、设边界、做判断，数字员工理解上下文、调工具、持续执行并交付完成的工作。 · 「超级个人 → 超级团队」：个人侧 LightClaw / QClaw 等，企业侧以 WorkBuddy 为入口，把人、Skill、连接器和管控面收成可复制的执行力。文中还区分狭义办公（个人生产力入口）与广义办公（真正改 CRM / ERP / MES 里的业务状态）。 手册中也对 WorkBuddy、CodeBuddy、ClawPro、ADP、SkillHub 等在企业中的定位做了详细解释和区分。 企业内部提效（方案册第 2 块） · WorkBuddy：用「业务大脑」拼全局判断；举例经鼎捷 MCP 操作真实 CRM/ERP/MES，强调联调验的是改状态而不只是能回答。上岗配置、财务/销售/交期等「以前 → 用了」对照页很多。安全侧自称工信部首批、信通院云端智能体基线首批、个保影响评估「二星级+」等。 · CodeBuddy：从 AI 辅助到 AI 原生；规约（SPEC）显性化；需求—编码—QA—交付链路由 Agent 打底、人审验收。 · 游戏研发 / 运营：策划与原型走 WorkBuddy，研发联调走 CodeBuddy；运营侧 DataBuddy 等数据 Agent。文称策划周期、Demo 周期明显缩短（具体倍数为自述）。 · SRE / 数据库：WorkBuddy × CloudQ 做故障定位；DatabaseClaw 称积累十万级工单经验、1600+ OpenAPI、十余款库、分级处置。 · 案例名：盛和、淘米（内容效率近 80%、覆盖约 80% 员工）、游族（47 模型、Agent 渗透率超 75%，并挂海外收入/毛利率变化）、游卡（MTTR 5–15 分钟、运维成本约 -40%）、搜狐畅游（P0 恢复缩短、重复运维 -70% 等）、花椒、紫龙、马帮等。 业务创新的几个章节 · SaaS Agent：订阅 → Credits → 按结果付费；界面操作走向 Skill/API。点名明略、熵简、慧算账、富通 JoinfClaw、蓝色光标、六度人和、光魔、天财商龙等。 · 内容生产：营销图视频、游戏美术（Miora / Ardot 等）；2D 序列帧等效率对比为文中自述。筷子、LiblibAI、必凡、玩心不止、紫讯、小影、懂球帝等。 · 社交陪伴 / 游戏内 AI / 客服：TokenHub + 音视频与语音链路；NPC 记忆与任务；客服/质检/外呼（天润融通、轻喜到家、智齿等）。延迟与解决率数字均为文中宣称。 · 生态共建：Buddy 应用、专家、Skill、连接器、硬件；自称桌面办公智能体头部、千万级安装、十万级生态资产、百余伙伴（文中自述）；智慧芽、蜜度、MasterGo 等出现在伙伴页。 完整报告，可以私信我获取（不是刻意引流哈，是因为没找到免费可公开访问的 PDF 地址，是一个付费下载地址，私信或留言我可以把 PDF 发给你）
Body · AI translation
This week, we continue with a complete study of this 78-page "2026 AI Agent Enterprise Solutions and Implementation Practices Handbook" jointly released by Tencent, Tencent Cloud, and WorkBuddy.
Today is Chapter 3, "Technological Innovation - The Technology Foundation from Model Training to Agent Applications." It splits the technology foundation that Agents need into two parts: on the training side, it uses an extremely elastic sandbox to improve Agentic RL post-training efficiency; on the application side, it uses Agent Runtime (sandbox + gateway + identity + observability + memory) to let digital employees run safely in production. The two case studies, StepFun and Kimi, bring "millions of images / tens of thousands of concurrent instances" and "millisecond startup and shutdown + fine-grained outbound network access" into large-scale real-world enterprise scenarios.
I uploaded the handbook to WorkBuddy: https://t.co/QeVenO5LDx
# 3.1 How Sandboxes Improve Post-Training Efficiency
The handbook argues that for general models to reach the next level, single-turn RLHF/RLVR is not enough; they must move to Agentic RL: let the model observe, act, and receive feedback in real multi-step tasks, then score by trajectory. This requires reproducible "real environments": fixing GitHub issues, buying tickets on a phone, browser operations, and so on. Sandboxes therefore change from auxiliary tools into core post-training infrastructure.
The three major challenges are described in very engineering-oriented terms: first, pulsed large-scale startup and shutdown; a typical step may instantly require starting thousands to tens of thousands of sandboxes (the text gives the example batch 512 × group 16 ≈ 8192), and if it is slow, GPUs run idle or training even fails; second, environments must expand from "can run code" to a full-scenario digital world (code / browser / phone / desktop / SWE, etc.); third, high-quality interaction trajectories and checkpoint snapshots must be retained to support replay, branching experiments, and full-link observability.
The corresponding solution points are: resource pooling and snapshot recovery, which the text says can deliver a usable sandbox in about 60ms; massive image distribution and on-demand loading, with the goal of starting about 160,000 instances of different images within about 1 minute; four mainstream types of sandboxes plus evaluation environments such as SWE / Android World / OS World; snapshots are used not only for crash recovery but also for reinforcement learning techniques such as Monte Carlo tree search that "return to the previous node and search in parallel."
Benchmark: StepFun StepFun uses Tencent Cloud sandboxes to handle million-level image storage and preheating, rapid startup of tens of thousands of concurrent instances, and connects to TokenHub for horizontal multi-model evaluation, linking the entire post-training chain from image building → preheating → environment simulation → trajectory synthesis → evaluation. The text says post-training efficiency increases by about 30%, while also raising GPU utilization. The related products named are Agent Sandbox, TCR, COS, TokenHub, and CCN.
# 3.2 Agent-Native Infrastructure (Application Side)
Scenarios range from Coding Agent, Deep Research, data processing and office automation, to GUI Agent (Browser / Computer / Mobile Use). The challenges shift into a production tone: behavior cannot be fully predicted, so hard isolation is needed to avoid deleting files or configuration changes spilling over; persistent Agents (harnesses such as OpenClaw / Hermes) need to be stateful 24/7, yet cannot really occupy full compute all the time, so they need on-demand pause/resume; access to external systems must be controllable and auditable, and keys must not enter the Agent's field of view.
English text is being prepared.
Benchmark: Kimi Moonshot's Kimi Agent has moved from long text to deliverable workflows; the text says Agent Swarm has up to about 300 sub-agents and more than 4,000 tool calls. Tencent Cloud provides: MicroVM strong isolation, millisecond-level startup and shutdown (the time to start a sandbox for complex tasks drops to the order of about 80ms), Agent-oriented storage (exclusive CBS / shared CFS, balancing throughput with human presentation and online collaboration), and fine-grained permissions and token control at the outbound gateway. The customer value statement is that it supports the stable growth of Agent clusters at the multi-million-user level.