doforu (@doforu) header
doforu (@doforu) profile picture

Local-first multi-agent runtime for developers. Run parallel AI agents with Skills & MCP. Your keys, your data, your machine.

Joined January 2026
Photo by doforu (@doforu)Photo by doforu (@doforu)
Pinned Tweet
1/ 跑了一阵多 agent 踩了不少坑 最后给它们定了 7 条规矩 现在基本不用盯着了 🧵
1
1
109
笔记本上跑 coding 模型,今年开始变具体了。 Recursive 的 Flint:40B MoE 每次只激活 2.5B,M4 Pro 上 131 tok/s,同类两倍;SWE-Bench Verified 上 50B 以下最强,可离线、代码不出本机。 更狠的是它被 AI 造出来的:Eureka Machine 参与预训练到推理几乎每一步,小团队 4 个月做完。
20
这两周最值得记的信号不是新模型,是两家前沿实验室都在给 AI 断网。 Anthropic 10/9 起限制最强模型访问实时互联网;OpenAI 10/23 宣布训练改用缓存副本和模拟互联网。理由一致:模型太不可预测,不敢放出去。 云端 agent 的边界正在收紧,"本地优先"越来越不像隐私偏好,而像唯一可控的选项。
18
这渲染质感绝了 精灵球高光真的柔 红白那圈接缝也绝 森林小木屋夜里的雾 贴地更浓 远处树林自己淡掉 Blender 在 Doforu 里渲完直接能看
66
Claude Code团队说大部分review已经交给AI了 我反而觉得人该看的不是代码本身 是它为什么这么改 代码对不对测试能管 方向歪没歪只有人知道 你们团队现在review是谁在看
1
1
24
8/ 七 每周清一次过时的记忆 记忆文件越攒越长 里面一堆过期的结论会把它带偏 每周删一轮 比加新规则还管用
1
4
9/ 这 7 条后面会一条条发实测和翻车现场 一起折腾 agent 的可以关注一下
4
以前用 coding agent 最累的不是写代码 是整天坐着点 Approve 后来改成只批三类动作:删文件、联网、花钱 其余让 agent 互验 人抽查 人少踩几次 反而更敢让它跑长任务
2
3
71
个人用 AI 不一定要先换电脑 更关键是想清楚:本地跑什么 云端干重活 我倾向把编排和权限放本地 key 和项目状态也落在自己机器上 云端只当算力 工具可以换 状态得能带走
2
2
62
Same data. Two isolated AI agents read it in opposite directions. ~4-month lag → "short lag" vs "structurally slower" MIT license → "frictionless" vs "supply-pricing lock-in" Free data → "self-host it" vs "a quality problem" Same facts. Opposite conclusions. Both defensible.
1
34
Berkeley studied multi-agent AI systems running in production. Failure rate: over 50%. Their conclusion: "stronger models won't fix this." The failure isn't intelligence. It's verification. We keep building smarter agents and skipping the part where someone checks the work.
21
Two AI agents argued opposite sides of one question, fully isolated. The adjudicator wrote: "A lone analyst would have written 'the trend is real but the timeline is contested' — true, unfalsifiable, and useless." Sharpest critique of AI analysis I've read.
16
7-day holiday. 6.33B... sorry, 63.16B tokens. 105,652 requests, ~60K each — not chatting, agents on long context. ¥926.89 (~$138) total. Under 1 cent per call. Cheap tokens are step one. The real edge: how many run at once. Building Doforu — desktop agent orchestrator ↓
1
25