Yicheng Wang

Building Runner, a local workspace where coding agents work together.

[Yicheng Wang @yicheng47]

@jiangkoumo_ https://t.co/gtwpAGZe26

自己写的orchestrator也天天在修这块东西，索性基本上状态现在都用hook来解决了。

[Yicheng Wang @yicheng47]

@DanKornas I’m the author of this repo. Thanks for the support.

[Yicheng Wang @yicheng47]

@sunsetsyntax I’m the author of the project, thanks for your support, you can dm me if you have some feature ideas

[Yicheng Wang @yicheng47]

@neerajjj6785 Still terminal, just like the TUI, and they still update a lot features on the TUI version. For tab management and orchestration with other TUI, I build my own tool https://t.co/gtwpAGZe26, you might want to take a look.

[Yicheng Wang @yicheng47]

Runner 0.13.0 is out. https://t.co/x7d1TEj0lm

You can now quit the app while your chats and crew missions keep running. Open Runner again to reconnect to the same sessions.

Crew messages and human requests still work with the app closed.

macOS and Windows, MIT licensed.

[Yicheng Wang @yicheng47]

@huacnlee 写完让 opus 5.5 review 下

[Yicheng Wang @yicheng47]

@ZacharyZhang computer-use 做所有的回归测试后，突然感觉轻松了很多，之前总感觉应该自己验收，现在完全放弃

[Yicheng Wang @yicheng47]

The key is the tooling. If you still only use terminal to manage, the ceiling is low. At least you should try something like herdr or cmux. I went all the way to build my own ai shed, check it out if you're interested https://t.co/ICpl8ZTQri

[nick @thecsguy]

if you don’t trust your agents, you will be slowed down.

build verification and validation tooling your agents can use, and benefit from being more hands off.

micro managing each thread is so Q1 2026 😄

[Yicheng Wang @yicheng47]

Runner now has a website: https://t.co/ICpl8ZTQri

The landing page was done by the following crew defined in Runner:

- coder (gpt-6.1-sol xhigh): implementataion
- reviewer (gpt-6-astra xhigh): code review
- qa (gpt-6.1-sol xhigh): uses the computer-use plugin for e2e testing

[Yicheng Wang @yicheng47]

@nini_webdev Zed is better.

[Yicheng Wang @yicheng47]

Just realized something. Workflow definition and agent workflow run are actually different things.

The workflow could be a DAG or anything, but which agent runs it isn't fixed. Most frameworks seem to start each node as a new thing. But agents have context, why not reuse them?

[Yicheng Wang @yicheng47]

@yifanxu_ephai 我最近一个两年的号也被封了，应该是最近得了甲流咨询了两句中文，再也不和claude讲中文了

[Yicheng Wang @yicheng47]

Lately, I’ve been polishing the UI components and refactoring parts of the implementation of my local agent orchestrator app, Runner (https://t.co/1mtzUjxU7T).

The app integrates deeply with agent hooks and takes an opinionated approach to how agents should work together locally, based on my own work experience. For example, it uses reusable role and crew templates, favors TUI over ACP, and uses an embedded CLI over MCP. I hope you like it.

The next step would probably be to decouple the terminal PTY processes from the app, making it more like a terminal multiplexer, but I’m still unsure whether to take that step.

Below is a demo of the latest UI. The entire project is written in Rust, with GPUI as the UI layer and Alacritty for terminal parsing. MIT Licensed.

[Yicheng Wang @yicheng47]

@xiaoduoya_ 字节内部的ai是很落后的，而且团队之间的差距非常大

[Yicheng Wang @yicheng47]

@Gray99999999 @lencx_ 这个价值还是有的，多个cli工具不太会做跨模型的，而且在gui上还是比herdr这种tui的要更友好一些

[Yicheng Wang @yicheng47]

@lencx_ https://t.co/gtwpAGZe26

能不能自荐个类似的东西，业余时间弄着玩的，套壳之上加了orchestator编排和多agent沟通的channel

[Yicheng Wang @yicheng47]

@huacnlee 非常同意，相当于在做项目的同时和Agent一起学习和探索，作为项目owner的价值主要在critical thinking 和 technical taste上

[Yicheng Wang @yicheng47]

@cristicrtu I've been building something very similar, rust + gpui to build a local agent orchestrator: https://t.co/gtwpAGZe26

Could you share the repo you're working on?