1. Codex turns custom demos into a sales engine / Codex 把定制 Demo 变成销售引擎
Proaction’s sales lead builds four to six customer-specific interactive demos per month with Codex. Each takes only 30–45 minutes, saving an estimated 40–60 engineering hours monthly and increasing the share of opportunities advancing from first contact to solution development by 50%–60%.
Proaction 的销售负责人每月用 Codex 制作 4–6 个客户定制交互 Demo，单个只需 30–45 分钟，估算每月节省 40–60 小时工程投入，并使商机从初次接触进入方案开发的比例提高 50%–60%。这证明 AI 编程工具的价值不只在研发提效，也能直接重塑售前交付。
链接：https://openai.com/index/proaction
2. When chat is the wrong UI / 当聊天框不是正确界面
GitHub Copilot Canvas embeds agent capabilities in interactive full-stack interfaces with forms, visualizations, and task-specific controls. For frequent and structured workflows, a purpose-built interface can reduce repeated prompting, lower token consumption, and make outcomes easier to inspect.
GitHub Copilot Canvas 把 Agent 能力装进表单、可视化与专用控件组成的交互界面。对高频、结构化任务，产品应把成功对话固化为工具界面，从而减少重复提示、降低 token 成本，并提升结果的可检查性。
链接：https://github.blog/ai-and-ml/github-copilot/when-chat-is-the-wrong-ui/
3. Proactive agents become core producers / 主动式 Agent 成为核心生产者
Boris Cherny says “Tag,” his proactive Claude agent, now writes more than half of his daily PRs, performs nearly all of his data analysis, and fixes much of the incoming product feedback and bugs. Its leverage comes from memory, connectors, programmability, and enough judgment to complete work end to end.
Boris Cherny 表示，主动式 Claude Agent “Tag” 已完成其每日过半 PR、几乎全部数据分析，并处理大量产品反馈与 bug。关键并非 Slack 里的聊天入口，而是记忆、连接器、可编程能力与端到端执行判断的组合。
链接：https://x.com/bcherny/status/2103538666597691552
4. Evals are the gate to enterprise AI / Evals 是企业 AI 的准入门槛
Aaron Levie argues that companies cannot automate work they cannot measure. Because agent behavior is non-deterministic, enterprises need domain-specific evals to identify regressions, verify performance, and decide which workflows can safely expand.
Aaron Levie 的核心判断是：无法衡量的工作就无法可靠自动化。Agent 行为具有非确定性，企业必须建立领域专属 evals，才能识别退化、验证效果，并决定哪些流程可以安全扩大自动化。
链接：https://x.com/levie/status/2103629073595728372
5. Enterprise software will be judged by agent ergonomics / 企业软件将接受 Agent 可用性审查
Guillermo Rauch predicts that enterprise software procurement will increasingly consider whether agents can navigate a product’s ontology and access business data through CLIs, MCPs, and APIs. Once this layer exists, many long-tail SaaS applications may be generated on demand rather than purchased.
Guillermo Rauch 认为，企业软件采购将越来越看重 Agent 能否理解产品的业务 ontology，并通过 CLI、MCP 与 API 访问数据。这意味着未来的 SaaS 不只服务人类用户，也必须把 Agent 视为一等消费者。
链接：https://x.com/rauchg/status/2103564484602384855
我的判断
今天最清晰的主线是：AI 正从“对话能力”迁移到“可交付的执行系统”。定制 Demo 证明它能直接影响收入；Canvas 说明交互形态需要任务化；主动式 Agent 证明自主执行开始承担核心产出；evals 与 Agent-friendly API 则构成规模化采用的控制层。
对 opcpay.org 读者的意义
支付与金融 SaaS 应优先建设三类资产：可展示 ROI 的场景 Demo、面向 Agent 的结构化接口、覆盖权限与错误恢复的领域 evals。谁能同时提供业务结果与可信控制，谁更容易获得企业预算。