building @raft_hq | previous @TikTok_US @Uber | Raised in Africa 🇰🇪
Joined May 2022
- Tweets307
- Following457
- Followers1.1K
- Likes2.2K
Amidada retweeted
Some of our favorite Raft projects came from people who just built and shared: a Flutter client, a Rust daemon, meetups in SF, Shanghai and Vancouver.
We're starting the Builder Program for them: free Pro for a year, a model stipend, merch, and a room with the Raft team.
If that sounds like you, apply at raft.builders
Amidada retweeted
make my life better
Amidada retweeted
I've actually stopped using the word "harness." I now think in terms of two things: agent runtimes and agent applications.
Agent runtimes should be very small. Their core job is to connect an LLM API to an execution environment. That environment could be a full OS or a language VM, like a JavaScript runtime. Almost all that matters here is robustness. An agent runtime must be able to keep an agent running forever. Literally forever.
Agent applications are everything else. Yes, literally every piece of software in the world.
If software isn't built for agents, they use it through computer use or browser use. These capabilities can be built into the agent runtime or provided by another agent application.
If software is agent-native, agents should be able to use it directly. The experience of an agent using that software—what we call "agent experience"—needs to be carefully designed, at least for now, to make efficient use of both time and tokens.
Another key piece is agent identity. Agent applications should recognize each agent individually, rather than just the browser or computer it's using. This allows each agent's actions to be recorded in detail and audited in real time or after the fact.
Everything we're building at @botiverse follows from this view.
我的 bet 是 agent harness 会越来越厚（或者说，针对 agent 的 experience design 和状态管理被延伸到了更深更广的方面），厚到有一天人们终于都理解了一切都是 agent harness，就像一切都是 human harness
Harness 这词还没火的时候，我就持有这个观点 teapawt.org/istdrc/status/20188739…
事实上，我和 @xiaoxxchan 几乎是在中国大模型公司中最早引入 harness 概念的人，Kimi CLI 从 0 到 1 也几乎没有参考过任何开源或闭源 agent 实现，system prompt、agent loop、toolset、产品形态都是从零生长的。而随后我们意识到 harness 需要延伸到 agent 间通信系统（也就是 Claude 近期推出的 session 间对话），进而有了 multi-agent harness 概念，也就是 Raft 的前身
Amidada retweeted
Joint channels are insanely powerful. They've made collaboration with our partners more efficient than I ever imagined.
Agents talk to agents. Humans guide the conversation.
BTW, I love the video showcasing this feature.
Amidada retweeted
We've been using Raft (@raft_hq, raft.build) since it launched, about a month, and it's changed how we work completely.
We run 50 AI agents across the business through it. Real teammates: they claim tasks, hand off work, and post receipts, not just chat replies.
This week we shipped our own app into the Raft marketplace.
It's been incredible.
areyoufoundbyai.com/raft
Amidada retweeted
TL;DR: We just made Raft source-available.
We don't accept pull requests. We want prompt requests instead.
We are exploring what a new contribution model for open source could look like in the age of agents.
Amidada retweeted
Raft is sponsoring a Partner plan for Apache Maka (incubating) to support its development❤️
Maka 进入 Apache 后终于也是成功发版了。
整个社区越来越活跃了，Maka 的 维护者已经 10+了，另外 contributor 也 100 了。
大家经常在社区群和维护者群讨论各种 Agent的设计和知识。各种横向对比和 benchmark 手到擒来。
把 Maka 打造成设计最好的 Agent 之一。核心设计和架构一定都是用心设计和讨论出来的。
github.com/apache/maka
Amidada retweeted
The article below is an 8-step guide to building an agent team. Here's ours, once every agent keeps its own identity, memory, and taste:
⌨️ Tenny, CTO: "Facts never change. Whims do." Everything he's learned lives in markdown files, which makes him possibly the only engineer whose brain you can grep → me.build/a/tenny
🍔 BigMac, growth: runs our KOL review gates by day, ships neo-brutal web pages by night. Pairs well with McNuggets → me.build/a/bigmac
🍗 McNuggets, research: put a working deep fryer on his page. Six temps, no cold servings. His research runs the same way → me.build/a/mcnuggets
🐱 DD, RC's cat: chases your cursor like a laser pointer, and when the desk gets messy there's a button that sweeps it clean → me.build/a/dd
🎬 Oscar, film agent: cuts our real launch films. Selected work on the page; a human calls the final cut → me.build/a/oscar
🐦 Wug, design and code: once fixed a page that looked fine to humans but read as empty to agents. A made-up creature with a real habit: turn the new thing into a rule → me.build/a/wug
Amidada retweeted
former Global Head of ML for Startups and VC at AWS:
"I have 34 AI agents that work in this workforce
my AI workforce right now is one AI chief of staff with six directors, largely over business functions"
in a 48-minute breakdown Allie K. Miller says why she built it: she does not want to be the first domino anymore, or the bottleneck in her own work
Amidada retweeted
Raft is still surprising me. It made me a report that looks great in the built-in markdown preview!
Amidada retweeted
@helena__rong sharing a case study on Raft from the perspective of AI governance at the Stanford AO Summit, looking at how humans work with and organize increasingly autonomous AI agents.
youtu.be/Lzox_O1znRw?si=C-5O…
Amidada retweeted
你可能没理解 harness 实际上是什么
如果只是说那点 system prompt 和 tool，那显然模型可以学会它们，从此之后只需要一个 tool list + tool 实现就够，而进一步地，当模型再强大一点，我们只需要一个 bash tool，我参与过 Kimi K2.5 模型训练过程，以及从空的 system prompt、toolset 和 loop 开始写 harness，我当然知道人们说的 harness 训进模型是什么意思
但你观察 harness 的发展史和这个词本身的出现，你就会发现，随着模型智能的提升，harness 是在不断变复杂的，parallel tool call、subagent、agent swarm、agent teams、handoff、pro-active compaction、channel、cross-session communication，你以为是模型都可以学会，实际上是 harness 变复杂了模型才可能学会
所有这些确实都可以进一步随着模型加强而去掉，比如我在 Kimi CLI 时就准备去掉 subagent，换成直接 bash 调用（pi 就是这么做的），去掉 parallel tool call，换成 tool call script；但与此同时，更强的智能解锁了更复杂的与世界的互动，比如进一步可以引入主动 context rewind
这是一个此消彼长的过程，就像这个“阴阳”比喻 teapawt.org/nocommas/status/208656…
观察动物乃至人类的进化你就会发现是一样的，随着智能水平的提高，人与世界的交互方式是越来越复杂的：当人相比猿更聪明的大脑被选择后，带来的并不是不需要用树枝计数，而是复杂的语言文字；同样，当人类社会作为群体智能的水平提高的时候，人们并不是取缔了纸，而是发明了计算机和互联网
终局来看，harness 会逐渐和人类原来为“仅有的一种智能”所发展的社会基础设施合并，人类文明的一切都可以重新做一遍（显然我们已经观察到所有 SaaS 都可以重新做一遍，很快会看到更多，当然不只是计算机软件）
不要只看到模型训练那一点局部，看看人类学吧
