- 发布时间
参考自动驾驶从 L1 到 L5 的分级方式，我们不讨论具体技术细节，只看 AI 最终能达到的效果与人机分工，把 AI Agent 的能力分为 AL1 到 AL5：从代码补全、对话问答、自主探索的协作者，到提出超越专家方案的引领者，再到完全不需要人参与的自主系统。
Multi-agent
全部文章
- workflow (14)
- linux (12)
- ai (12)
- opencode (12)
- tools (10)
- multi-agent (10)
- arch-linux (9)
- productivity (9)
- ai-agent (9)
- ai-coding (9)
- claude-code (8)
- ide (7)
- neovim (6)
- wayland (5)
- ikanban (5)
- window-manager (4)
- tiling (4)
- sway (4)
- coding (4)
- plugins (4)
- subagents (4)
- codex (4)
- vibe-coding (3)
- mcp (3)
- cli (3)
- journey (3)
- vim (3)
- dotfiles (3)
- reflection (3)
- configuration (3)
- development (3)
- surfingkeys (3)
- budget (3)
- deepseek (3)
- dynamic-workflows (3)
- rust (3)
- i3 (2)
- river (2)
- skills (2)
- vscode (2)
- editor (2)
- copilot (2)
- setup (2)
- guide (2)
- agent (2)
- introduction (2)
- chat (2)
- automation (2)
- misc (2)
- hyprland (2)
- browser (2)
- blog (2)
- research (2)
- verification (2)
- pricing (2)
- kimi (2)
- video-generation (2)
- superpowers (2)
- ai-workflow (2)
- development-tools (2)
- deepseek-harness (2)
- vibe-kanban (2)
- parallel-workflow (2)
- task-scheduling (2)
- qwen (2)
- 2025 (1)
- niri (1)
- scrollable-tiling (1)
- screen-sharing (1)
- webrtc (1)
- desktop-environment (1)
- x11 (1)
- keyboard-first (1)
- jetbrains (1)
- autohotkey (1)
- windows (1)
- docker-compose (1)
- experience (1)
- daily-use (1)
- security (1)
- openness (1)
- life (1)
- summary (1)
- philosophy (1)
- claude (1)
- specialized-agents (1)
- development-workflow (1)
- github (1)
- godot (1)
- game (1)
- dev (1)
- laptop (1)
- i3wm (1)
- nvidia (1)
- wlroots (1)
- nixos (1)
- declarative (1)
- reproducible (1)
- deploy (1)
- oss (1)
- bot (1)
- js (1)
- typst (1)
- typesetting (1)
- document (1)
- template (1)
- academic (1)
- keymaps (1)
- autonomy-levels (1)
- human-in-the-loop (1)
- chrome-devtools-mcp (1)
- token (1)
- openrouter (1)
- tts (1)
- python (1)
- manim (1)
- cache-hit (1)
- tmux (1)
- academic-research (1)
- cordis (1)
- multi-session (1)
- agent-runtime (1)
- ultracode (1)
- orchestrator (1)
- fzf (1)
- fuzzy-finding (1)
- ade (1)
- ai4paper (1)
- human-ability (1)
- danmu (1)
- streaming (1)
- real-time (1)
- bilibili (1)
- md2video (1)
- agent-first (1)
- markdown (1)
- cui (1)
- harness (1)
- tool-use (1)
- gesture-recognition (1)
- magic-mouse (1)
- multi-touch (1)
- multi-provider (1)
- glm (1)
- keyboard-navigation (1)
- gui (1)
- evdev (1)
- 发布时间
长时间运行的 AI agents 可以生成超过我们吸收能力的代码与研究成果。这些输出扩展了我们的生产能力，却不会自动转化成我们的知识、判断与能力。- 发布时间
Claude 账号被封之后，备用方案重新成为主力。本文介绍 GPT、OpenCode、iKanban 和基于 skill 的 AI Workflow 如何让工作继续，也讨论它与 dynamic workflows 的差距，以及为什么编排能力更适合成为可移植的插件，而不是绑定在 runtime 中。- 发布时间
过去我们手动拆任务、一个个启动 agent。如今 Claude Code 的 dynamic workflows（ultracode 模式）会自己写出编排脚本、调度一整支 subagent 队伍——本文通过构建 iread RSS 阅读器的三个真实工作流来展示，并把 agent 之间的异步通信留作下一个待解问题。- 发布时间
在 OpenCode 之上构建工具和工作流几个月之后，我们把日常使用切回了 Claude Code。原因是具体的：GPT 调整账号安全策略之后模型性价比发生变化、把 Claude 走 OpenCode 带来的封号风险、prefix cache 失配带来的周配额损耗，以及一个并没有带来质变的 browser 包装实验。