# What does it really take to create autonomous AI coworkers?
**作者**: Shiv
**日期**: 2026-04-18T15:55:10.000Z
**来源**: [https://x.com/shivsakhuja/status/2045531591926329474](https://x.com/shivsakhuja/status/2045531591926329474)
---
Ever since I discovered OpenClaw, I've been obsessed with building AI agent coworkers for my startup.
I've tried tons of setups over the past few months using a mix of OpenClaw / Claude Code / Claude Agent SDK and our own product @GooseworksAI.
The dream is to have a team of agents that do useful work autonomously around the clock without constant supervision.
I wouldn't say we're in the dream state yet, but it's getting close and I've learned a lot about what it takes.
I think the reason that I (and many others like me) are obsessed with this idea, is because it's so clear that the agents like Claude Code are already good enough to do all the actual work. The bottleneck is all just orchestration.
Here's what I've learned about the pieces you need to orchestrate a coworker:
0. Environment
When we're talking about AI coworkers, they need an environment to run free. They need their own computer.
You can't run them locally otherwise, they can't work when your computer is off, and its too risky to give them full access to your personal machine.
If you're using OpenClaw, this could be a Mac Mini or an EC2 instance.
For @GooseworksAI, we use cloud sandboxes powered by @e2b.
1. Context
Your agent is only as good as the context it has.
The thing with context is – it requires a continuous stream, not a one-time dump.
I do a few things:
- I do frequent 1:1s with my agent: https://x.com/shivsakhuja/status/2038711292073320734
- My agent has a email address, a phone number and a Slack account, so I'm always texting it stuff - links I find interesting, notes, ideas, etc
- I have a cron job to sync meeting notes
- I have a personal CRM for my agent (updates automatically from calendar + email + meeting notes + linkedin) so it knows about people I'm connected to
- I have given it accounts so it can search through Notion + Linear
- I have signed up my agent to some newsletters (using @agentmail)
- I have an agent consume a content diet similar to what I consume (twitter mostly, but will probably add some podcasts soon via @ListenNotes)
- I have a heartbeat for the agent to read through recent notes and ask me questions about important context in Slack. It's pretty easy for me to respond when I'm prompted with a message.
I haven't tried it yet, but from what I hear @garrytan's gbrain is a pretty good system to stream context too.
2. Persistence
We’re using a filesystem. We dump / stream tons of context in there for each agent. This does very well for the most part. But there is a limit beyond which it's not sufficient.
The agent has a memory folder where it stores important info automatically. The agent also has a tool to search through it’s chat history (past sessions)
I’ve also found SQLite DB to be super helpful for structured query-able persistence. For example, we have an SEO agent, and that needs to keep track of past work, de-dupe, etc before we run a strategy, push to our CMS, email for backlinks, etc. So currently that’s using a SQLite DB in the filesystem for persistence.
There’s probably a very good case to be made for adding a knowledge graph / dedicated memory layer though we haven’t tried this yet.
Our agents also read / write from Linear to manage tasks, which is very helpful.
3. Skills — atomic units of execution
A skill is a repeatable task packaged into a prompt + script so your agent can execute it consistently.
We made some of these skills public at http://skills.gooseworks.ai/ but we have tons more in an internal github repo.
If nothing else, just turning everything into skills makes a massive difference. I'm making skills for everything now.
This makes my agent good at the things I'm good at.
An orchestrator skill has also been pretty useful – this is a parent skill that knows how to orchestrate related sub-skills.
4. Automations and/or a heartbeat — when does it run?
Still very much figuring this out, but my current mental framework is:
- If it's a clear repeatable task → scheduled automation (cron). Pull metrics every Monday, scrape inspiration accounts every Sunday, etc.
- If the agent needs to figure out what to do dynamically → heartbeat. The agent wakes up on an interval, checks what's changed, decides what to act on.
I’m still tinkering with heartbeats but getting the agent to make good prioritization decisions when it wakes up is harder than it sounds.
For many things, I’m just running the skills / orchestrator manually because I don’t trust the agent enough yet.
5. Tools and access
The agent needs to actually do things — hit APIs, send messages, scrape the web, etc
I use a mix of MCP servers and direct API calls wrapped in skills.
Haven't found a "one system to rule them all" — it's whatever works for the task at hand.
6. Communication channels — two-way
A channel where you can message the agent and it can message you. We use Slack primarily, but I can also email or text the agent.
The key is two-way — not just notifications, but actual back-and-forth. We also integrated WhatsApp, iMessage and Telegram, but mostly I just use Slack.
An advantage of Slack is team visibility (the agents are shared, not personal)
Email is important too because a lot of the web works over email (notifications, newsletters, etc).
7. Feedback loop
This is the piece that most people skip but seems super important to me.
The agent does work → you review it → your feedback gets fed back into the system (updated skills, adjusted prompts, new rules).
The agent self-reflects → proposes learnings / updates to its own system
If you have this and it works, the agent’s usefulness / success should (theoretically) compound over time. Without it, the agent only improves when you do dev work.
8. View layer
There are 2 challenges here.
Viewing files in the filesystem. We solved this with our own product (has a baked in filesystem and file viewer), but others may solve this with Obsidian remote vaults or something.
Agent activity dashboard: We have this in Gooseworks, but I found this to be pretty tough in OpenClaw.
Custom dashboards: I need custom dashboards for specific agents. Like I need a custom dashboard for my Outbound sales agent. I don’t have a great solution here right now though we're experimenting with one.
We use Slack alerts but that can get very chaotic if you’re relying on the agent to do a bigger scope of work.
Not sure what the best option is. I suspect it's to give each agent a DB + Dashboard and allow it to customize.
9. One agent or multiple?
I wrote a post about this: https://x.com/shivsakhuja/status/2035176670286786744
the TLDR is 1 agent if you don't mind mixing context + tool access. Multi-agent if you need strict separation.
The challenge with multi-agent is how these agents communicate if they are silo-ed.
I have not yet solved this, but I'd love to talk to anyone who has.
10. Tying it all together - the systems mindset
In theory, OpenClaw / Gooseworks exposes all these pieces but you still need to have an engineering mindset and stitch them all together, and engineer the system in the right way.
This is often the hardest part.
For example, let's say I want my agent to find me leads by scraping LinkedIn posts for some keyword. Sounds like an easy problem to solve. I wire up my agent to an Apify actor and run an automation, right?
But not really. Because the agent will just send me the same leads every day. No deduping happening. Now if I have the LLM dedupe, it's incredibly inefficient. So I need to make sure that my Apify scraper is ONLY checking the last 24 hours. But the Apify scraper doesn't have a way to filter by timestamp, so now what? Also, the same people tend to post about the same topics a lot.
This can be solved by sending the leads to a CRM and making sure there's a way to track the outreach status there, but the point I'm making is that these are actually engineering problems and require a systems mindset to solve. The example above is probably easier than most real systems.
What I'm curious about:
I imagine a lot of people are figuring this stuff out right now, so I'd love to trade notes:
- What are you trying to get your agent coworker to do? What's the dream scenario 6 months out?
- How do you structure your skills and feedback loops?
- What's your context layer? just a filesystem or anything else?
- Has anyone gotten a good heartbeat-based system working?
- What does your agent check for when it "wakes up"?
Feel free to comment and / or DM me. You can also find me on LinkedIn at https://linkedin.com/in/shivsakhuja
## 相关链接
- [Shiv](https://x.com/shivsakhuja)
- [@shivsakhuja](https://x.com/shivsakhuja)
- [15K](https://x.com/shivsakhuja/status/2045531591926329474/analytics)
- [@GooseworksAI](https://x.com/@GooseworksAI)
- [@GooseworksAI](https://x.com/@GooseworksAI)
- [@e2b](https://x.com/@e2b)
- [https://x.com/shivsakhuja/status/2038711292073320734](https://x.com/shivsakhuja/status/2038711292073320734)
- [@agentmail](https://x.com/@agentmail)
- [@ListenNotes](https://x.com/@ListenNotes)
- [@garrytan](https://x.com/@garrytan)
- [http://skills.gooseworks.ai/](http://skills.gooseworks.ai/)
- [https://x.com/shivsakhuja/status/2035176670286786744](https://x.com/shivsakhuja/status/2035176670286786744)
- [https://linkedin.com/in/shivsakhuja](https://linkedin.com/in/shivsakhuja)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [11:55 PM · Apr 18, 2026](https://x.com/shivsakhuja/status/2045531591926329474)
- [15.8K Views](https://x.com/shivsakhuja/status/2045531591926329474/analytics)
- [View quotes](https://x.com/shivsakhuja/status/2045531591926329474/quotes)
---
*导出时间: 2026/4/19 19:15:31*
---
## 中文翻译
# 打造自主AI同事，到底需要什么？
**作者**: Shiv
**日期**: 2026-04-18T15:55:10.000Z
**来源**: [https://x.com/shivsakhuja/status/2045531591926329474](https://x.com/shivsakhuja/status/2045531591926329474)
---
自从我发现 OpenClaw 以来，我就一直痴迷于为我的创业公司构建 AI 智能体同事。
过去几个月里，我尝试了大量的配置方案，混合使用了 OpenClaw / Claude Code / Claude Agent SDK 以及我们自己的产品 @GooseworksAI。
梦想是拥有一支智能体团队，能够全天候自主完成有用的工作，而不需要持续的监督。
我不能说我们已经达到了梦想状态，但已经越来越接近了，而且我学到了很多关于这需要什么的知识。
我认为我（以及很多和我类似的人）之所以痴迷于这个想法，是因为很明显，像 Claude Code 这样的智能体已经足够优秀，可以完成所有实际工作。瓶颈只在于编排调度。
以下是我学到的关于编排一个“同事”所需的各个组成部分：
0. 环境
当我们谈论 AI 同事时，它们需要一个可以自由运行的环境。它们需要自己的电脑。
否则你无法在本地运行它们——你的电脑关机时它们就无法工作，而且让它们完全访问你的个人机器风险太大。
如果你使用 OpenClaw，环境可以是一台 Mac Mini 或一个 EC2 实例。
对于 @GooseworksAI，我们使用由 @e2b 提供支持的云端沙箱。
1. 上下文
你的智能体的能力取决于它所拥有的上下文。
上下文的关键在于——它需要持续的信息流，而不是一次性的灌入。
我做了以下几件事：
- 我经常和我的智能体进行 1:1 沟通： https://x.com/shivsakhuja/status/2038711292073320734
- 我的智能体有一个邮箱地址、一个电话号码和一个 Slack 账号，所以我随时可以给它发消息——我觉得有趣的链接、笔记、想法等等
- 我设置了一个 cron 任务来同步会议记录
- 我为我的智能体建立了一个个人 CRM（自动从日历 + 邮件 + 会议记录 + LinkedIn 更新），这样它就了解我的人脉关系
- 我给它开通了账号，让它可以搜索 Notion + Linear
- 我给我的智能体订阅了一些新闻通讯（使用 @agentmail）
- 我让智能体接收和我类似的内容（主要是 Twitter，但可能很快会通过 @ListenNotes 添加一些播客）
- 我为智能体设置了一个心跳机制，让它阅读最近的笔记，并就重要的上下文在 Slack 上向我提问。当它主动发消息提示我时，我回复起来非常方便。
我还没尝试过，但据我所知 @garrytan 的 gbrain 是一个相当不错的上下文流式传输系统。
2. 持久化
我们使用文件系统。我们为每个智能体存入/流式传输大量上下文。这在大多数情况下效果很好。但存在一个上限，超过这个上限就不够用了。
智能体有一个记忆文件夹，它会自动在里面存储重要信息。智能体还有一个工具，可以搜索它的聊天记录（过去的会话）。
我还发现 SQLite 数据库对结构化的、可查询的持久化非常有帮助。例如，我们有一个 SEO 智能体，它需要跟踪过往工作、去重等，然后才能执行策略、推送到我们的 CMS、发送外链邮件等。所以目前它使用文件系统中的 SQLite 数据库进行持久化。
添加知识图谱/专用记忆层可能是一个非常好的方案，不过我们还没有尝试过。
我们的智能体还会读写 Linear 来管理任务，这非常有帮助。
3. 技能——原子化的执行单元
技能是一个可重复的任务，被打包成提示词 + 脚本，以便你的智能体能够一致地执行它。
我们在 http://skills.gooseworks.ai/ 上公开了一些这样的技能，但我们在内部 GitHub 仓库里还有更多。
即便别的都不做，仅仅把所有东西都转化为技能就会带来巨大的差别。我现在正在为所有事情制作技能。
这让我的智能体擅长我所擅长的事情。
一个编排器技能也非常有用——这是一个父技能，知道如何编排相关的子技能。
4. 自动化和/或心跳机制——什么时候运行？
这方面我还在摸索中，但我目前的思维框架是：
- 如果是一个明确可重复的任务 → 定时自动化（cron）。比如每周一拉取指标，每周日抓取灵感账号，等等。
- 如果智能体需要动态决定要做什么 → 心跳机制。智能体按间隔醒来，检查有什么变化，然后决定处理什么。
我仍在调试心跳机制，但让智能体在醒来时做出良好的优先级决策，比听起来要难。
对于很多事情，我只是手动运行技能/编排器，因为我还不够信任智能体。
5. 工具和访问权限
智能体需要真正地做事——调用 API、发送消息、抓取网页等等。
我混合使用了 MCP 服务器和封装在技能中的直接 API 调用。
还没找到“一统天下的唯一系统”——什么适合手头的任务就用什么。
6. 沟通渠道——双向的
一个你既能给智能体发消息、它也能给你发消息的渠道。我们主要使用 Slack，但我也可以通过邮件或短信联系智能体。
关键是双向——不仅仅是通知，而是真正的来回交流。我们还集成了 WhatsApp、iMessage 和 Telegram，但我大部分时候只用 Slack。
Slack 的一个优势是团队可见性（智能体是共享的，而不是个人的）
邮件也很重要，因为网络上的很多事情都是通过邮件进行的（通知、新闻通讯等）。
7. 反馈循环
这是大多数人都会跳过的部分，但对我来说似乎极其重要。
智能体完成工作 → 你进行审查 → 你的反馈被回馈到系统中（更新技能、调整提示词、新增规则）。
智能体进行自我反思 → 提出对自身系统的学习/更新建议
如果你有这个机制并且它有效，那么智能体的实用性/成功率应该（理论上）随着时间不断复利增长。没有它，智能体只有在你做开发工作时才会改进。
8. 视图层
这里有两个挑战。
在文件系统中查看文件。我们用自己的产品解决了这个问题（内置了文件系统和文件查看器），但其他人可能会用 Obsidian remote vaults 或类似工具解决。
智能体活动仪表盘：我们在 Gooseworks 中有这个功能，但我发现要在 OpenClaw 中做到这一点相当困难。
自定义仪表盘：我需要针对特定智能体的自定义仪表盘。比如我的外呼销售智能体就需要一个自定义仪表盘。我目前还没有很好的解决方案，不过我们正在试验一种方法。
我们使用 Slack 提醒，但如果你依赖智能体完成更大范围的工作，这可能会变得非常混乱。
不确定最佳方案是什么。我猜是给每个智能体配备一个数据库 + 仪表盘，并允许它自定义。
9. 单智能体还是多智能体？
我写过一篇关于这个的文章：https://x.com/shivsakhuja/status/2035176670286786744
简单来说：如果你不介意上下文和工具访问权限混在一起，就用单智能体。如果需要严格分离，就用多智能体。
多智能体的挑战在于，这些智能体被隔离后如何相互通信。
我还没有解决这个问题，但很乐意和已经解决的人交流。
10. 把一切串联起来——系统思维
理论上，OpenClaw / Gooseworks 已经暴露了所有这些组件，但你仍然需要具备工程思维，把它们串联起来，并以正确的方式构建整个系统。
这往往是最难的部分。
举个例子，假设我想让我的智能体通过抓取 LinkedIn 上某个关键词的帖子来为我寻找潜在客户。听起来是个容易解决的问题。我把智能体连接到一个 Apify actor 并运行自动化，对吧？
但其实不然。因为智能体每天都会给我发送同样的潜在客户，没有去重。如果让 LLM 去重，效率会极低。所以我需要确保我的 Apify 抓取器只检查过去 24 小时内的内容。但 Apify 抓取器没有按时间戳过滤的功能，那怎么办？而且，同样的人往往反复发布相同主题的内容。
这可以通过把潜在客户发送到 CRM 并确保在 CRM 中有跟踪触达状态的方法来解决，但我想说的是，这些实际上是工程问题，需要系统思维才能解决。上面的例子可能比大多数真实系统都要简单。
我好奇的是：
我想现在有很多人正在摸索这些东西，所以我非常乐意交流心得：
- 你想让你的智能体同事做什么？6 个月后的理想场景是什么？
- 你如何构建你的技能和反馈循环？
- 你的上下文层是什么？只是一个文件系统还是有其他东西？
- 有人成功搭建过基于心跳（heartbeat）的系统吗？
- 你的智能体在“醒来”时会检查什么？
欢迎评论和/或私信我。你也可以在 LinkedIn 上找到我：https://linkedin.com/in/shivsakhuja
## 相关链接
- [Shiv](https://x.com/shivsakhuja)
- [@shivsakhuja](https://x.com/shivsakhuja)
- [15K](https://x.com/shivsakhuja/status/2045531591926329474/analytics)
- [@GooseworksAI](https://x.com/@GooseworksAI)
- [@GooseworksAI](https://x.com/@GooseworksAI)
- [@e2b](https://x.com/@e2b)
- [https://x.com/shivsakhuja/status/2038711292073320734](https://x.com/shivsakhuja/status/2038711292073320734)
- [@agentmail](https://x.com/@agentmail)
- [@ListenNotes](https://x.com/@ListenNotes)
- [@garrytan](https://x.com/@garrytan)
- [http://skills.gooseworks.ai/](http://skills.gooseworks.ai/)
- [https://x.com/shivsakhuja/status/2035176670286786744](https://x.com/shivsakhuja/status/2035176670286786744)
- [https://linkedin.com/in/shivsakhuja](https://linkedin.com/in/shivsakhuja)
- [升级到 Premium](https://x.com/i/premium_sign_up)
- [2026年4月18日 晚上11:55](https://x.com/shivsakhuja/status/2045531591926329474)
- [15.8K 次浏览](https://x.com/shivsakhuja/status/2045531591926329474/analytics)
- [查看引用](https://x.com/shivsakhuja/status/2045531591926329474/quotes)
---
*导出时间： 2026/4/19 19:15:31*
