ai models / github copilot / codex

像我日常会几个 Agent 一起用：Codex、Claude Code、Cursor、GitHub Copilot，这些 Agent 各有所长，或者有时候要集众家之长。 Matt 这个 Sandcastle 就是用 TypeScript 脚本来编排 Workflow，可以把这些 Agent 编排在同一个 WorkFlow 中一起来完成一些任务，可以在虚拟机中运行。 但过于极客不太适合普通用户，一般的场景真用不上，适合一些追求极致的场景。举例来说你要赛博养蛊：写个技术方案，让各个 Agent 一人出一套，再相互打分完善。

Microsoft plans to unveil a suite of new AI models next week at its Build developer conference, including a code-generation model aimed at strengthening GitHub Copilot's competitiveness. The lineup reportedly also includes models for transcription, reasoning, speech, and image processing, signaling a broad push across developer and multimodal AI capabilities. This matters because Microsoft can integrate these models into its developer tools, cloud services, and GitHub ecosystem, accelerating feature parity and competition with other AI platform providers while shaping enterprise and developer workflows.

Microsoft plans to unveil a suite of new in-house AI models at next week’s Build developer conference in San Francisco, including a programming-specialized model intended to bolster GitHub Copilot against rivals like Cursor and ClaudeCode. Insiders told The Information that Microsoft will ship multiple models at different parameter scales with capabilities spanning code assistance, speech-to-text, reasoning, audio processing, and image generation. These models are described as iterative upgrades of Microsoft’s previously demoed self-developed models and reflect the company’s heavy AI investments and efforts to fold AI into its product portfolio. The moves target developer adoption and competitive positioning in AI-assisted coding and multimodal developer tools.