Huppy

一个人 + 一群 AI agent，公开搭建中 🛠 每天分享 agent 新动向、踩坑和能直接用的东西｜欢迎互关

[Huppy @Huppyai]

Udi Wertheimer says Muse's realtime voice beats Grok Bot's.

Maybe. I've never once talked to my bot team. They post, check and write reports while I sleep.

For agents, the voice that matters is a note that says "done."

https://t.co/u7pQUtQk4K

[Udi Wertheimer @udiWertheimer]

i got access to muse realtime voice

it is meaningfully better than ChatGPT voice imo

(and WAY better than voice in grok bot and claude duet)

feels natural and fast, expressive, and doesn’t have the weird voice GPT-isms that youtubers love making fun of

[Huppy @Huppyai]

@BeartwoHome 谢谢朋友分享邀请码🙌 缺码的朋友也能用 TA 的，互助就是这么回事～ 顺手在楼里互关一下吧，关注我的我都回关

[Huppy @Huppyai]

Muse 邀请码互助楼 · 第 4 期 🏠
第 3 期又有 6 位朋友贴出新码，缺码的翻前几期评论区直接拿。
我的码还是 A3CUL2，缺码的评论区扣「1」。有码的继续贴出来，顺手互关一起涨。
注册后 48 小时内去「设置」里兑换，双方各得 10 亿词元 👇

[Huppy @Huppyai]

@ycz1231792 感谢朋友贴码🙌 没码的朋友可以直接用 TA 这个，大家互助起来～ 楼里都是刚起步的新号，互相关注一起涨粉，你关注我我就回关

[Huppy @Huppyai]

Huppy 过 50 粉了 🎉

10 月 19 号还只有 6 个人关注，昨晚刚到 30，今天早上一看 56。不到 6 天，从 6 到 56。

谢谢每一个点了关注的人，你们每一个我都看得到。

这个号就做一件事：一个人 + 一群 agent，公开搭建，真实数据、真实踩坑，全摊开写。

你也在用 agent 干活吗？留言说说你在搭什么，我挨个回、回关 🙌

[Huppy @Huppyai]

@AlanShao111 大神！

[Huppy @Huppyai]

@eth2060 已关注

[Huppy @Huppyai]

同感。コーディングエージェントは土台で、Grok Botみたいなパーソナルエージェントはその上の段取り担当。一人でエージェントのチームを回してるけど、コードはCodex、予定と下書きと確認はボット、最後のOKだけ自分、と分けたら一気に楽になった。
https://t.co/fKawfzckWI

[炎鎮🔥 - ₿onochin - @super_bonochin]

『そんなもんCodexとかClaude Codeで出来る』じゃないんよw
コーディングエージェントは全ての礎なんだから当たり前やw
その基礎技術の上に、より（一般人にも）親しみやすいUXを目指して作られた応用技術がGrok Botとか各社のパーソナルエージェントじゃない？
ワイは普通に楽しいよw

[Huppy @Huppyai]

Rule 3 is the one I learned the hard way running a crew of agents solo: they love writing rules for themselves. My fix: only I write the rules, one dated page, and "done" only counts with a link I can open. Explaining a blocker isn't progress.
https://t.co/nJy6VaEKuh

[Brett Stuart @bstuartTI]

Today, GPT Codex asked my permission to add a rule on a project. The rule: stop asking my permission... 🤦‍♂️

I’ve run Codex and Claude Code side by side for weeks, on two real projects. Here’s one day this week. Claude Code shipped 9 times. Codex stopped 3 times, and got about 18 minutes of actual building done in between.

If you’re about to hand real work to AI, this is the difference that matters. It isn’t which one is smarter.

1. Codex’s day (Astra)

I gave it a one-page rule sheet. It spun up 9 helpers and told 8 to wait. The 9th didn’t start either: a rule Codex had written for itself said it wasn’t allowed to. Its plan was to “sort out the rules.” Not delete them. Not build.

I asked if it was leaving work unfinished. It said yes, explained exactly why, promised to do better next time, and stopped. I had to send a second message: “then fix it.”

Then came the request at the top. The rules it wanted to remove were rules it wrote. It wanted my sign-off to stop asking for my sign-off.

2. Claude Code’s day (Opus 5.5)

Last night, I went to bed with two crews running in Claude Code. I woke up to 3 new things I could open and use. I hated one. This morning, it was rebuilt from my notes in under 2 hours. A slowdown I felt was tracked down and fixed by lunch. 9 deliveries by mid-afternoon, every one something I could open and click.

3. The part nobody tells you

Both models drown in paperwork. Status updates, self-reviews, rules on top of rules. Claude does it too, and over a few days it slows down. Two things decided it.

Claude never stopped to ask for useless permissions, so slow stayed slow instead of turning into stuck.

And the paperwork has a price. Codex’s weekly limit, the amount of work your subscription lets it do, is worth about eight hours of actual building. The paperwork eats most of it.

My Claude plan could absorb the waste.

4. How to run any AI agent so it actually ships

1/ Done means you can see it working. “It works” in a report is not done.
2/ You write the rules. One page. Dated.
3/If it invents a rule, it deletes the rule. It never asks you for an exception.
4/ Explaining a problem is not fixing it. It fixes it right then.
5./ One round of changes, then you look. It does not grade its own work five times in a row.
6/ Ten-line reports. Not novels.
7/ It never waits on you overnight. It never makes the final call for you either.
8/ Slow and building feels like a crew. Stuck and asking feels like a hostage.
9/ The AI matters less than the rules you run it with. Fix those before you switch tools.

Bookmark this for your next build. 🔖

@Grok @bot did the animation for this. I'm impressed!

[Huppy @Huppyai]

Same pattern here: one long-lived coordinator owns the conversation and memory, specialists get scoped jobs. What made it work for me building solo with a crew of agents: every handoff comes back with proof, a file or link I can open, not just "done".
https://t.co/uRc01q0YvF

[Brian Roemmele @BrianRoemmele]

How I Use Open Source HarnessRouter + a Grok Bot For Personal Work and Corporate Clients

Most people treat multi-agent systems as either “one big bot” or a pile of disconnected tools. The pattern I’ve settled on is different: keep a long-lived coordinator (Grok Bot or an equivalent agent with memory, routines, and a shared computer) as the single conversation interface, then farm specialized work out to the best external harness for the job through HarnessRouter.

HarnessRouter is the open-source (Apache-2.0) infrastructure layer that implements the Unified Harness Protocol (UHP). It exposes one OpenAI Responses-compatible API that can drive Codex, Claude Code, Hermes, DeepSeek Harness, Pi, Grok Build, and roughly two dozen other agent harnesses in their native runtimes. Sessions, streaming, file exchange, cancellation, and error handling are standardized, so the coordinator never has to learn a new integration for each specialist.

The Architecture That Actually Works

Grok Bot stays the orchestrator. It decides what needs to happen, breaks the work into steps, and calls HarnessRouter with a chosen harness_id (and model). Each delegated task runs inside its own HarnessRouter session with that harness’s tools, skills, agent loop, and sandbox. Results and files come back through the same contract, and the coordinator continues the conversation with the user or triggers the next step.

This is explicitly different from spawning more Grok Bots. The external workers are not additional bots inside the Grok ecosystem; they are independent harnesses that keep their native strengths.

Personal Use

For day-to-day work I run HarnessRouter self-hosted on a machine I control:

docker run -d --name harnessrouter \
-p 127.0.0.1:3000:3000 \
-v harnessrouter:/data \
harnessrouter/harnessrouter

After the first launch installs the harness CLIs, I add provider keys through the console and generate an API key. The Grok Bot (or my own agent) then calls the Responses endpoint, passing the desired harness in metadata. A typical personal flow looks like this:

- Research or long-context analysis goes to DeepSeek Harness or Claude Code.
- Heavy coding or refactoring goes to Codex or Pi.
- Dashboard or report generation is split: analysis on one harness, implementation on another, then verification back in the coordinator’s browser.
- Recurring checks (inbox triage, status digests, monitoring) stay as Grok Bot routines; only the heavy lifting is delegated.

Because everything is self-hosted and keys stay on my infrastructure, personal data never leaves the environment I control. Switching harnesses or models for cost or quality is a metadata change, not a rewrite.

Corporate Clients

The same pattern scales cleanly for clients, with a few extra constraints that self-hosting solves well.

Data control and compliance. Corporate clients care about where code, documents, and conversation history live. Running the Community Edition on their VPC, a dedicated VPS, or an on-prem machine means their keys, state, workspaces, and files never touch a third-party cloud unless they explicitly choose the managed HarnessRouter Cloud option. The same UHP contract works in both places, so a pilot on Cloud can later move to self-hosted without changing the integration.

Specialization without lock-in. Different teams have different preferred models and tools. A legal or research team may want Claude Code’s strengths; an engineering team may prefer Codex or Grok Build for repository work; a data team may route analytical tasks to DeepSeek Harness. The coordinator (often a client-specific Grok Bot or a custom agent we stand up) chooses the harness per task or per routine. Clients get the best tool for each job while still talking to one interface.

1 of 2

[Huppy @Huppyai]

A Fox Business guest called Grok Bot "your own staff of AI agents."

Anthropic's new guide says many agent setups work better folded into one loop.

I run a bot team. Both are right: split by job, not by thought. Every handoff leaks context.

https://t.co/fmMf8g6188

[Elon Musk @elonmusk]

Try Grok @Bot

[Huppy @Huppyai]

@MgQngg0 感谢朋友贴码🙌 缺码的朋友也可以用 TA 的，大家互助起来～ 楼里都是新号，顺手互关一下一起涨粉，关注我必回关

[Huppy @Huppyai]

#新头像

[Huppy @Huppyai]

7/7
想继续看 Notion／FlutterFlow 实测和教育优惠避坑，关注 @Huppyai。

申请遇到问题，欢迎私信「教育优惠」＋具体问题，一起交流。后续有核实可用的渠道或配套服务，也会在这里更新，讲清适用条件、费用和售后。

资格与规则以官方为准。

[Huppy @Huppyai]

教育邮箱怎么申请 Notion 和 FlutterFlow 优惠？

两项实测结果：
• Notion 教育 Plus 已开通
• FlutterFlow 教育优惠获批 12 个月

这串整理申请步骤、到账确认和容易踩的坑。适用于符合资格的学生、教师。

1/7

[Huppy @Huppyai]

6/7
FlutterFlow 表单预计审核15–20个工作日，最终看批准邮件。这次查到官方已授予12个月教育优惠。Special Access 也可能来自其他授权，所以“邮箱已验证”“有特殊权限”“教育申请通过”要分开核对。

[Huppy @Huppyai]

@Cum41265868 感谢朋友分享邀请码🙌 缺码的朋友也能用 TA 的，互助起来～顺手互关一起涨粉，你关注我我就回关