1. Real-SWE：公开榜单不再足以评估编码 Agent
Real-SWE: Re-evaluating coding models on private enterprise repositories
Real-SWE moves software-engineering evaluation from public repositories to real private enterprise codebases, directly testing the effects of training-data contamination and benchmark overfitting.
Real-SWE 把评测从公开仓库移到真实企业私有代码库，直接触及训练数据污染与公开基准失真的问题。企业采购 Agent 时，应把重点转向自有代码、权限边界和交付验证。
链接：https://withspecific.com/benchmarks/real-swe
2. HydraFusion：多模型编排开始成为产品能力
HydraFusion approaches frontier quality through multi-model orchestration
GitHub's research preview selects models by task. In controlled offline evaluations, its selective coding workflow matched or exceeded the tested Opus 5 baseline while reducing estimated cost.
GitHub 的研究预览按任务选择不同模型；在受控离线评测中，其选择性编码工作流达到或超过被测 Opus 5 基线，同时降低估算成本。模型路由正从后台工程优化变成用户可感知的产品能力。
链接：https://github.blog/ai-and-ml/github-copilot/project-hydrafusion-frontier-quality-via-multi-model-orchestration/
3. Devin + Astra：Agent 不只要写代码，还要证明写对了
Cognition uses GPT-6 Astra to let Devin test its own work
Cognition is using Astra to have Devin run tests and present verification evidence, aiming to reduce the need for engineers to inspect every generated line manually.
Cognition 让 Astra 驱动 Devin 执行测试并展示验证结果，目标是减少工程师逐行审查代码的负担。编码 Agent 的竞争正在从“能写”转向“能证明写对了”。
链接：https://openai.com/index/cognition-devin-testing-with-astra
4. Plugin evals 与真实任务评测成为基础设施
Anthropic adds plugin evals while builders question pass/fail benchmarks
Anthropic introduced claude plugin eval init so developers can verify whether skills still work after model releases. Builders also warn that strict hidden tests can reject answers that are more sensible than expected outputs.
Anthropic 推出 plugin evals，用于检查模型升级后 skills 是否仍能正常工作；与此同时，一线开发者提醒，过严的 hidden tests 可能拒绝比标准答案更合理的结果。评测需要从单一 pass/fail 走向真实任务、过程证据和人工判断的组合。
链接：https://x.com/trq212/status/2098531560643539440
5. Agent economy 需要微支付与组织记忆
Coinbase: agentic finance needs native accounts, micropayments, and organizational memory
Brian Armstrong says roughly 76% of agent e-commerce payments observed by Coinbase are below $0.30. Coinbase also gives each engineering repository a “brain” that records incidents, controls, experiments, and accepted or rejected pull requests.
Brian Armstrong 表示，Coinbase 观察到的 agent 电商支付约 76% 低于 0.30 美元。Coinbase 还为每个工程仓库建立“brain”，持续记录事故、控制要求、实验及 PR 接受/拒绝历史。微支付解决机器间交易，组织记忆解决持续纠错，两者都是 Agent 成为经济主体的底层条件。
链接：https://www.youtube.com/watch?v=uLDK4l_-gUE
我的判断
今天最强的主线不是某个新模型，而是 Agent 的评测、验证、路由和记忆正在合并为可信执行系统。公开 benchmark 只能回答“平均能力如何”，企业真正需要回答的是：在我的数据、权限、成本与验收规则下，它能否稳定交付。
对 opcpay.org 读者的意义
支付和金融天然要求高权限、强审计与明确责任边界，因此会比一般 SaaS 更早需要 Agent control plane。内容上应把 Real-SWE、HydraFusion、验证闭环与此前 SWE-2 成本研究并成“质量—成本—验证”系列；产品上应优先关注可审计执行、回滚、模型路由和低额机器支付。