2026-10-22 AI / SaaS 情报简报
1. OpenAI extends prompt caching to 12 hours / OpenAI 将 Prompt Cache 延长至 12 小时
OpenAI has extended API prompt caching to 12 hours. For logged-in Codex and ChatGPT Work users, returning to an older conversation after several hours should still produce a cache hit. OpenAI also corrected a usage-metering mismatch affecting roughly 40,000 accounts and says impacted users will receive credits.
OpenAI 把 API prompt caching 延长至 12 小时，长时间运行或中断后恢复的 Agent 因而能减少重复上下文成本。约 4 万账户的计量偏差也提醒产品方：usage ledger、对账与补偿机制必须与 Agent 能力同步建设。
2. Vercel opens a durable runtime layer for agents / Vercel 开放 Agent 持久运行层
Vercel is opening the durability infrastructure behind Eve and Workflows as a runtime-agnostic layer for agent harnesses such as Pi and OpenCode. It can run locally, on Vercel, or in self-hosted environments, adding persistence and observability to long-running agents.
Vercel 正把 Eve 和 Workflows 背后的基础设施开放为 runtime-agnostic agent layer，并支持本地、云端和自托管。长期 Agent 的基础能力正在从“会调用工具”升级为可暂停、恢复、观测和迁移。
3. Split-agent supervision outperforms loose subagents / 职责分离的 Agent 工作流优于松散协作
Peter Steinberger reports that Astra handled planning, invoked Codex CLI in isolated worktrees for implementation, and used a second agent to challenge the work. He packaged the planner-executor-critic pattern as an OpenClaw plugin called “smart goal.”
该实践把 planner、executor、critic 的责任显式分离，并用隔离 worktree 控制实现边界。真正提升可靠性的不是 Agent 数量，而是角色契约、验收标准和独立质疑机制。
4. Domain-specific autonomous systems become a durable product form / 领域自主系统成为稳定产品形态
Box CEO Aaron Levie argues that task- and domain-specific autonomous systems are moving the market beyond simple LLM wrappers. Their advantage combines stronger tool use, better domain judgment, and mature infrastructure such as sandboxes, filesystems, and data pipelines.
领域 Agent 的护城河不是 Prompt，而是工具执行、专业判断与 harness infrastructure 的组合。企业采用还取决于人类监督、权限边界和组织变革，单点模型能力不足以完成替代。
5. Agent orchestration inherits distributed-systems problems / Agent 编排重演分布式系统问题
Aditya Agarwal compares inference-time scaling, subagents, and agent swarms with the evolution from CPU threads to multicore processors and distributed systems. Coordination, isolation, scheduling, observability, and failure recovery are therefore becoming first-class product concerns.
Agent swarm 并不会消除复杂性，而是把复杂性转移到协调、隔离、调度、可观测性和故障恢复。未来的 Agent 平台更像 distributed systems control plane，而不只是聊天界面。
我的判断
今天的信号共同指向一个结构变化：模型能力正在成为底层供给，Agent 产品的差异化上移到 durable execution。谁能让任务安全地持续运行、恢复状态、隔离变更、解释成本并处理失败，谁才拥有生产级入口。
对 opcpay.org 读者的意义
Agent 支付不能只提供交易 API。它需要与执行状态绑定：为每个任务记录身份、预算、缓存与工具成本、授权节点、失败重试和最终结果。最小可行方案是建立 task-level usage ledger，再叠加预算策略与异常暂停能力。