Ask HN: AI 智能体 vs. 网关 vs. 框架
1 分•作者: snthpy•7 天前
您好，
我想听听大家对 AI Agent 生态系统中不同组件的看法。我发现目前的术语非常令人困惑，因为在研究可用选项时，我并不总是清楚我实际会得到什么。
很多东西都被称为 Agent，但这个术语似乎很难定义，因为它通常指代重叠的功能。对我来说，Agents 目前似乎由以下组件构成：
## Harnesses (控制框架)
* 在 LLM 周围添加 UI 和系统指令，并且可能通过记忆、工具调用等方式对其进行增强。
* 例如：Claude Code, Code, Gemini CLI, pi.dev, ...
## Gateways (网关)
* 这些将 Agents 连接到您选择的通信工具，例如 Whatsapp、Telegram、Slack、Discord 等。
* 例如：OpenClaw 和 Nanoclaw
## Sandboxes (沙盒)
* 隔离的环境，Agents 可以在其中以有限或可审计的功能运行。
* 例如，可以是物理沙盒，如独立的 Mac Mini，也可以是 docker-agent、agent-sandbox、localsandbox 等。
然后，我们进入 *Agents*：
## Agents (智能体)
* 能够行使一定自主权的 AI 系统，即在没有人工干预的情况下具有一定程度的自主行为。
* 这些通常结合了以下组件：
```
- **LLM**: 提供智能/推理/决策能力
- **Harness**: 引导行为并实现超越当前上下文窗口限制的长期行为一致性
- **Gateway**: 与人类和/或其他 Agents 交互
- **Sandbox**: 旨在将 Agent 的能力限制在一组经过批准的安全操作中，以便 Agent 自主执行。
```
我的问题是，当我查看可用选项时，通常不清楚我得到了哪些组件：
```
- https://github.com/openclaw/openclaw: Typescript，OG，一站式解决方案，最容易安装，但显然有点像“科学怪人”，而且很复杂
- https://github.com/qwibitai/nanoclaw: Typescript，越来越受欢迎，极简主义，实现似乎依赖于 Claude Code 来处理安装和配置
- https://github.com/sipeed/picoclaw: Golang
- https://github.com/zeroclaw-labs/zeroclaw: Rust
- https://github.com/nullclaw/nullclaw: Zig
- https://nemoclaw.bot/: 由 Nvidia 推出，声称是“基于”开源，但我找不到代码库
```
我真正希望的是能够组装自己的 Sandbox + Gateway + Harness + LLM 堆栈。有什么想法/建议吗？
查看原文
Hi,<p>I'd like to get everyone's take on the different components in the AI Agents ecoysystem. I find the current terminology quite confusing as it's not always obvious what I'll be actually getting when I examine the available options.<p>A lot of things get called Agents, but that term seems hard to define as it often refers to overlapping functionality. To me, agents seem to be currently composed of the following components:<p>## Harnesses<p>- Adds UI and system instructions around an LLM and may also augment it with tools such as memory, tool calls, etc...<p>- Examples are Claude Code, Code, Gemini CLI, pi.dev, ...<p>## Gateways<p>- These connect agents to your communication tool of choice, e.g. Whatsapp, Telegram, Slack, Discord, ...<p>- Examples are OpenClaw and Nanoclaw<p>## Sandboxes<p>- Isolated environments where Agents can run with limited or auditable capabilities<p>- Examples can be from physical ones like separate Mac Minis to docker-agent, agent-sandbox, localsandbox, ...<p>This then takes us to *Agents*:<p>## Agents<p>- AI Systems capable of exercising some agency, i.e. non-zero amounts of autonomous behaviour without a human-in-the-loop.<p>- These usually combine the following components:<p><pre><code> - **LLM**: to provide the intelligence/reasoning/decision making
- **Harness**: to steer the behaviour and enable long-term coherence of actions beyond the limits of current context windows
- **Gateway**: to interact with humans and/or other agents
- **Sandbox**: that aims to limit the capabilities of the Agent to a sanctioned set of safe operations for the agent to execute autonomously.
</code></pre>
My issue is that when I look at the available options, it's generally not clear to me which components I am getting:<p><pre><code> - https://github.com/openclaw/openclaw: Typescript, The OG, one-stop solution, easiest to install but apparently a bit of a Frankenmonster and complex
- https://github.com/qwibitai/nanoclaw: Typescript, gaining popularity, minimalist, implementation seems reliant on Claude Code to handle install and configuration
- https://github.com/sipeed/picoclaw: Golang
- https://github.com/zeroclaw-labs/zeroclaw: Rust
- https://github.com/nullclaw/nullclaw: Zig
- https://nemoclaw.bot/: By Nvidia, claims to be open source "based", but I couldn't find a repo
</code></pre>
What I would really like is to be able to assemble my own stack of Sandbox + Gateway + Harness + LLM. Thoughts/recommendations?