WorkBuddy and Codex push AI competition toward task delivery and reliability
A 36Kr analysis, reposted with permission from the 亿欧智库 WeChat column, argues that AI applications are shifting from answering questions to delivering checkable results, with reliable execution becoming the adoption barrier. It uses Tencent's WorkBuddy, where users supply an Excel or CSV file plus metrics and time range and the product reads data, computes statistics and generates charts or reports, and OpenAI's Codex, whose 2025 launch handled code reading, command execution and tests before OpenAI described knowledge-work uses such as reports, spreadsheets, slides and research analysis in June 2026, as parallel examples. The piece also surveys fine-tuning, RAG, workflows and agents as intersecting routes, cites Dify's Human Input node and Salesforce Agentforce's separation of deterministic logic from model reasoning, and points to an ICML 2026 evaluation of 15 models that found capability gains yielded only modest reliability improvement.
The analysis frames product capability as task completion and repeatability rather than answer quality, which is a useful lens for judging agent products, but it introduces no new release or measured result of its own.
让AI回答“怎么做一份经营分析”,和让它读取销售表、计算指标、生成图表并交出报告,是两个不同层次的产品要求。WorkBuddy和Codex的实践表明,大模型应用开始承担更多实际操作,用户对产品的判断也从回答是否精彩,延伸到任务是否完成。
大模型应用正在进入任务交付的竞争,可靠执行成为落地门槛。一份可验收的成果,需要模型理解目标,也需要最新信息、工具操作、业务规则和结果检查共同参与。模型能力增强,正在重新划分这些环节的职责。
这也解释了微调、RAG、工作流和Agent为何持续交叉:它们分别改善模型能力、任务信息和执行过程。理解路线演变,应先看系统怎样把一次回答组织成一项可以完成、复核和交付的工作。
01 WorkBuddy和Codex让AI直接处理文件并交付成果
WorkBuddy和Codex值得放在一起观察,因为它们把任务执行能力带进了用户真实使用的工作环境。用户交出目标、资料和允许操作的范围,系统读取信息、调用工具并生成文件,成果再回到原来的工作过程。产品承担的职责因此从提供答案,延伸到组织一次实际执行。
#Tencent#OpenAI#Anthropic#WorkBuddy#Codex#Dify#Salesforce#Agentforce#Berkeley BAIR#MCP