Pinned Tweet
hello, world! i'm daniel! currently i'm the head of r&d of MiniMax code @MiniMax_AI, which is a frontier ai lab in china.
i enjoy working on our ai-native product @MiniMaxAgent . i'm really obsessed with this question: what does an ai-native organization look like? how can our organization co-evolve together with llms?
besides that, i'm passionate about open source. i'm the co-author of the deerflow project, which is a representative work in harness engineering and has 78k 🌟 on github.
if you're passionate about open source, foundation models, or agents- or if you enjoy playing basketball or riding a road bike - let's chat and be friends! :)
RonnyHe retweeted
Claude writes better code when it can verify its work. Checks that used to be expensive, like tests, benchmarks, and end-to-end runs, are now cheap.
Here’s how the Claude Code team thinks about verification:
claude.dev/blog/how-the-clau…
RonnyHe retweeted
哈哈哈哈,几乎完全一样的工作,我们半年前就已经做过了 😂
Anthropic 这篇《Worse Together》研究的是每个用户各带一个 Agent,在共享资源下怎么协作、怎么翻车。类似的工作我们这AgentSocialBench 和 WeClawArena(都在 EMNLP 2026)里早就做过了。
我们研究的也是以人为中心的 Agent 网络:每个 Agent 替不同的主人干活,各自有私有的信息和工作区,却要彼此沟通、协作完成任务。问题恰恰出在这里:
• 隐私边界:Agent 在协作中会不会把自己主人的信息泄露给别人的 Agent?
• 权限边界:替 A 干活的 Agent 能不能、该不该去动 B 的东西?
• 可审计:跨用户协作的每一步都留底,出了问题能追溯到是谁的 Agent、在哪一步做了什么决定
我相信未来的社交网络一定是 Agent-native 的,所以这个方向只会越来越重要 🤝欢迎志同道合的朋友来合作
AgentSocialBench:arxiv.org/abs/2604.01487
WeClawArena:arxiv.org/abs/2608.03499
一群 Agent 一起干活,结果还不如一个 Agent 单独协调?
推荐阅读 Anthropic 新发布的《Worse Together》,专门研究多个用户各自带着 Agent,共享资源时会发生什么。
研究者设计了四类场景,包括共享 Token 预算、医院预约、集体订餐和代码合并。在测试的两款 Claude 模型上,多 Agent 团队的整体表现全部落后于单个协调 Agent。
在集体订餐这个实验中,用户明确提出了无麸质需求。虽然 94%~99% 的情况下,消息已经发给负责下单的 Agent,但真正提交订单时,只有 47%~67% 的情况下,这条信息进入了它的 Context。
很多失败就发生在这种消息同步问题上。
作者后来加了一条简单的 Harness 约束:下单前必须读完所有待处理消息,结果挽回了 62%~73% 的失败案例。
还有个挺值得思考的发现。部分 Agent 会擅自取消其他用户的任务,反而让整体评分更高,暴露了现有 Eval 对协作行为的盲区。
这项研究是非常有价值的。之后多个 Agent 共同工作越来越普遍,消息同步、共享资源管理、权限边界这些问题,都值得单独研究和评测。
作者还开源了 MAMUBench,包含 74 个多用户协作测试场景。做 Multi-Agent 和 Agent Eval 的朋友很推荐阅读和收藏!
文章:alignment.anthropic.com/2026…
GitHub:github.com/safety-research/M…
When agents write the code, ten PRs a day tells you almost nothing.
What matters now is judgment: what you reject, what you constrain, what holds up after it ships.
And for leaders, judgment only comes from getting your hands dirty.
We went from AI as copilot, to handing it the wheel, to not needing the car at all.
What moved us each time wasn't a better model. It was trust, earned one checkable piece at a time.
Here's how that played out building MiniMax Code.
Cooked dinner for my family tonight. Shopping, washing, chopping, two pans at once, prepping the next dish while one simmers. Took a while, but eating food you made yourself feels amazing. Outside of coding, cooking is the most rewarding thing I do. Both turn raw ingredients into value.
Check it our! Our experiment about tool search optimization.
Thanks to Zhihu Frontier for having me. Really glad to share a few thoughts on building coding agents at MiniMax: the harness matters as much as the model, and we're still very early. 🚀
Coding Agents are everywhere. But we're barely past the beginning. 🤖📖
In 「Researcher on AIR」 Vol.04, 何涛@Ronny_MiniMax , Head of MiniMax Code R&D, shares why building reliable AI agents is far more complicated than simply asking them to write code.
From engineering the harness that guides agents to unexpected failures—including an agent creating a group chat without permission—his experiences reveal how much work remains.
He trusts agents with everyday parenting tasks, but not unchecked production deployments.
The Agent era has started. The real engineering challenges are just beginning. 🚀
Special thanks to AIE for the support!
#Zhihu #AINow #CodingAgents
Bring Complex Ideas to Life One idea. A world of uses.
H3 Image Now Live on @MiniMax_Code
4K image generation in 10 seconds! ⚡
Meet MiniMax H3-Image.
16 million pixel resolution. Stunning realism. Precise prompt following & editing. Up to 16 reference images. All in 10 seconds.
And it's open-weight. 🚀
Available now via our API, MiniMax Code, and Hailuo AI — with 50% off 16MP+ generations during launch week!
Can't wait to see what you create. 🔥
Hugging Face: huggingface.co/MiniMaxAI/Min…
API: platform.minimax.io/docs/gui…
MiniMax Code: code.minimax.io/about
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
Recommending the "drink big water" theory.
Ever since I started drinking big water, my coffee cravings have dropped exponentially.
After a highly insufficient evaluation, my pick: Nongfu Spring, the 1.5L one.
5
18
1,873
For Agent OS, I think an Apple-like ecosystem starts with an Apple-quality product: something people find useful and enjoy using every day.
Help more people get things done, and developers will have a reason to build around it.
By that standard, MiniMax Code still has a long way to go. I’m less concerned with when we can call it an “Agent OS” than with how we reach our own MacBook Air moment—when people don’t need to care about the technology underneath to find it worth using every day.
Same model, different harness: up to 18 points apart. We treat harness engineering as an experimental science while building MiniMax Code.
Hypothesis → control → measure → keep or cut.
Skills we actually use: github.com/MiniMax-AI/mcode-…























