AI Agent周报：OpenAI Codex扩展至日常工作

In today’s issue:

• OpenAI ships Codex for everyday work
• Cursor releases the Cursor SDK
• Mistral launches Workflows orchestration
• DAIR.AI guide to building LLM knowledge bases
• Agentic Harness Engineering paper drops
• Cursor 3.2 multitask lands
• Claude Code adds push notifications
• Qwen open-sources Qwen-Scope SAEs
• AISI evaluates GPT-5.5 cyber capabilities
• AgenticQwen-30B-A3B closes tool-use gap

And all the top AI dev news, papers, and tools.

Top Stories

Codex for Everyday Work

OpenAI extended Codex from a coding agent into a general-purpose work agent. Users now pick a role (finance, data science, marketing, ops, research), connect the apps they actually use, and get suggested prompts that wire Codex into docs, slides, sheets, research, and planning across ChatGPT.

• Role-based onboarding: Codex ships preset roles for non-engineering teams, with per-role prompt suggestions and connector recommendations so a marketing or finance user can run a useful agent on day one without designing their own harness.
• Sheets, slides, and docs: The update adds materially better spreadsheet and slide generation plus cleaner doc workflows, pushing Codex into the same surface as enterprise copilots like Workspace and Microsoft 365 agents.
• 20% faster computer use: Codex’s computer-use agent runs 20% faster on the same tasks, narrowing the latency gap that has held browser and desktop automation back from being a daily-driver capability.
• Same agent everywhere: OpenAI is positioning a single Codex runtime across coding, research, and operations, so a Pro or Business user gets one agent that scales from “fix this PR” to “build a Q2 finance review.”

Codex for Work | Announcement

Read more