# Fact: your agent should not need a skill for everything.
**作者**: Zaid
**日期**: 2026-05-16T07:29:26.000Z
**来源**: [https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711](https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711)
---
Your agent should not need a skill for everything.
If your agent needs a skill for everything, it is not ready.
I have spent enough time building agents to believe this strongly:
It should not need a plugin, skill, MCP server, custom script, workflow file, or manual setup for every basic thing you expect it to do.
Skills are useful.
Extensions are useful.
Custom workflows are useful.
But if an agent needs a separate add-on before it can read a file, check git state, run a command, manage memory, schedule a task, fetch a URL, or ask for approval, then the core agent is not ready yet.
That is one of the reasons we built Mercury Agent the way we did.
Mercury comes with batteries included.
Not as a marketing phrase.
As a product principle.
Out of the box, Mercury includes built-in tools for filesystem work, shell commands, git operations, web fetching, scheduling, skills, system budget checks, messaging, and sub-agent delegation. The docs list tools like read_file, write_file, edit_file, run_command, git_status, git_diff, git_add, git_commit, git_push, fetch_url, schedule_task, budget_status, and delegate_task. (mercury.cosmicstack.org)
That matters because agents are not normal apps.
A normal app waits for a user to click a button.
An agent reasons through a task, decides which tools to use, and acts across your environment.
So the question is not only:
“Can I extend this agent?”
The better question is:
“What can this agent safely do before I start assembling extra parts?”
Mercury’s answer is simple:
Enough to be useful.
Not enough to be reckless.
That balance is important.
We do not believe an agent should be powerless by default.
But we also do not believe it should silently touch your system.
Mercury is permission-first.
The GitHub README describes Mercury as permission-hardened, with folder-level read/write scoping, pending approval flow, shell blocklists, and session-level permission modes like Ask Me and Allow All. It also states that Mercury asks before it acts. (GitHub)
That is the line we care about.
Agents should be capable.
But capability without boundaries becomes anxiety.
You should not feel scared every time your agent wants to run a command.
You should not wonder whether it changed a file without telling you.
You should not need to inspect every step because the system gave it too much freedom too early.
A good agent should earn trust through visible, controlled action.
That is why Mercury tracks what matters.
You can check status.
You can check progress.
You can inspect memory.
You can inspect background tasks.
You can stop agents.
You can halt the system.
You can reset when needed.
The docs include commands for /status, /progress, /tools, /skills, /memory, /tasks, and /budget, plus emergency controls like /halt, /stop, and /reset. (mercury.cosmicstack.org)
This is not decoration.
This is the control layer.
If an agent is going to run for long tasks, work in the background, use tools, and operate near your codebase, then tracking is not optional.
It is basic safety.
Mercury also supports coding workflows directly.
It has /code plan for analysis and options without coding, /code execute for step-by-step implementation, /code agent <task> for delegating coding work to a background sub-agent, and workspace commands for opening directories, refreshing git state, staging files, committing changes, and undoing file edits. (mercury.cosmicstack.org)
This is where I think many agent products are going to converge.
The first generation of coding AI was autocomplete.
The second generation was chat.
The third generation is not just “more autonomous coding.”
It is controlled delegation.
You will not only ask AI to write code.
You will ask agents to investigate bugs, review diffs, inspect files, summarize repo state, run tests, compare options, prepare plans, and work on isolated subtasks in parallel.
For that to work, the agent needs built-in judgment and built-in tools.
Not a blank shell waiting for you to bolt on every capability manually.
Mercury still supports skills.
That part matters.
Skills are great when you want specialization.
A skill can encode a repeatable workflow, a team convention, a tool-specific process, or a domain-specific behavior.
But skills should extend the agent.
They should not compensate for an empty core.
The core agent should already understand how to operate.
It should know how to read.
How to inspect.
How to ask.
How to track.
How to remember.
How to stop.
How to operate within limits.
Mercury’s Second Brain is part of that.
The README describes it as persistent structured memory using SQLite and FTS5 full-text search, with memory types, auto-extraction, conflict resolution, and consolidation. (GitHub)
That is important because agents should not be stateless workers forever.
If you keep repeating your preferences, project context, habits, goals, and constraints, the system is wasting your time.
An agent should remember what matters.
But memory also needs boundaries.
That is why Mercury’s memory is local and inspectable through commands like /memory. (mercury.cosmicstack.org)
The same philosophy applies to tokens.
Agentic systems can burn tokens quickly.
Long context.
Tool calls.
File scans.
Repeated planning.
Background tasks.
Sub-agents.
If the agent has no budget awareness, the user pays for chaos.
Mercury includes daily budget enforcement, auto-concise behavior after a threshold, and /budget commands to check, reset, set, or override usage. (GitHub)
Again, this is not a side feature.
It is part of making agents usable every day.
A real agent needs memory.
A real agent needs tools.
A real agent needs permissions.
A real agent needs progress visibility.
A real agent needs a stop button.
A real agent needs budget control.
A real agent needs to be useful before you install anything extra.
That is what we mean by batteries included.
Not bloated.
Not reckless.
Not closed.
Just capable enough to serve its purpose from the first run.
Then, when you need more, you extend it.
That is the model I believe in.
Core capabilities built in.
Specialized workflows through skills.
User control everywhere.
Mercury is still early, but this is the direction we are building toward:
An agent that is not scared of real work.
An agent that does not need a separate skill for every basic action.
An agent that remembers what matters.
An agent that asks before it acts.
An agent that can run from the CLI or Telegram.
An agent that can work in the background, delegate to sub-agents, and still stay under your rules.
That is the future I want.
Not an empty chatbot.
Not a reckless operator.
Not a plugin maze.
A capable agent with boundaries.
Mercury Agent is open source.
GitHub:
https://github.com/cosmicstack-labs/mercury-agent
## 相关链接
- [Zaid](https://x.com/Ctrl_Alt_Zaid)
- [@Ctrl_Alt_Zaid](https://x.com/Ctrl_Alt_Zaid)
- [18K](https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711/analytics)
- [mercury.cosmicstack.org](https://mercury.cosmicstack.org/docs/reference/built-in-tools)
- [GitHub](https://github.com/cosmicstack-labs/mercury-agent)
- [mercury.cosmicstack.org](https://mercury.cosmicstack.org/docs/cli-commands/in-chat-commands)
- [mercury.cosmicstack.org](https://mercury.cosmicstack.org/docs/cli-commands/in-chat-commands)
- [GitHub](https://github.com/cosmicstack-labs/mercury-agent)
- [mercury.cosmicstack.org](https://mercury.cosmicstack.org/docs/cli-commands/in-chat-commands)
- [GitHub](https://github.com/cosmicstack-labs/mercury-agent)
- [https://github.com/cosmicstack-labs/mercury-agent](https://github.com/cosmicstack-labs/mercury-agent)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [3:29 PM · May 16, 2026](https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711)
- [18.8K Views](https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711/analytics)
- [View quotes](https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711/quotes)
---
*导出时间: 2026/5/17 10:19:28*
---
## 中文翻译
# 事实：你的智能体不应该每件事都需要一个技能
**作者**: Zaid
**日期**: 2026-05-16T07:29:26.000Z
**来源**: [https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711](https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711)
---
你的智能体不应该每件事都需要一个技能。
如果你的智能体每做一件事都需要一个技能，那它还没有准备好。
我已经花了足够多的时间来构建智能体，从而坚信这一点：
它不应该为了每个你期望它完成的基本任务，都需要一个插件、技能、MCP 服务器、自定义脚本、工作流文件或手动设置。
技能是有用的。
扩展是有用的。
自定义工作流是有用的。
但是，如果一个智能体在读取文件、检查 git 状态、运行命令、管理记忆、安排任务、获取 URL 或请求批准之前，都需要一个单独的附加组件，那么这个核心智能体就还没有准备好。
这就是我们以这种方式构建 Mercury Agent 的原因之一。
Mercury 开箱即用（自带电池）。
这不是一句营销口号。
这是一个产品原则。
开箱即用，Mercury 包含了用于文件系统操作、Shell 命令、Git 操作、Web 获取、任务调度、技能、系统预算检查、消息传递和子智能体委派的内置工具。文档列出了诸如 read_file、write_file、edit_file、run_command、git_status、git_diff、git_add、git_commit、git_push、fetch_url、schedule_task、budget_status 和 delegate_task 等工具。(mercury.cosmicstack.org)
这一点很重要，因为智能体不是普通的应用程序。
普通的应用程序等待用户点击按钮。
智能体会推理任务，决定使用哪些工具，并跨越你的环境采取行动。
所以问题不仅仅是：
“我能扩展这个智能体吗？”
更好的问题是：
“在我开始组装额外部件之前，这个智能体能安全地做些什么？”
Mercury 的答案很简单：
足以发挥作用。
不足以鲁莽行事。
这种平衡很重要。
我们不认为智能体应该默认是无力的。
但我们也不认为它应该悄无声息地触碰你的系统。
Mercury 遵循“权限优先”原则。
GitHub README 将 Mercury 描述为经过权限加固的，具有文件夹级别的读/写作用域、待批准流程、Shell 阻止列表以及如“询问我”和“全部允许”等会话级权限模式。它还指出 Mercury 在行动之前会先询问。(GitHub)
这就是我们要关注的那条界线。
智能体应该是有能力的。
但如果没有边界的能力就会变成焦虑。
你不应该在每次智能体想要运行命令时都感到害怕。
你不应该怀疑它是否在未告知你的情况下更改了文件。
你不应该因为系统过早地给予了它过多的自由，而去检查每一步操作。
一个好的智能体应该通过可见的、受控的行动来赢得信任。
这就是 Mercury 跟踪关键信息的原因。
你可以检查状态。
你可以检查进度。
你可以检查记忆。
你可以检查后台任务。
你可以停止智能体。
你可以暂停系统。
你可以在需要时重置。
文档包含了 /status、/progress、/tools、/skills、/memory、/tasks 和 /budget 的命令，以及 /halt、/stop 和 /reset 等紧急控制命令。(mercury.cosmicstack.org)
这不是装饰。
这是控制层。
如果一个智能体要运行长任务、在后台工作、使用工具并在你的代码库附近操作，那么跟踪就不是可选项。
这是基本的安全保障。
Mercury 还直接支持编码工作流。
它拥有 /code plan 用于分析和选项而不涉及编码，/code execute 用于逐步实施，/code agent <task> 用于将编码工作委托给后台子智能体，以及用于打开目录、刷新 git 状态、暂存文件、提交更改和撤销文件编辑的工作区命令。(mercury.cosmicstack.org)
这就是我认为许多智能体产品将会趋同的地方。
第一代编码 AI 是自动补全。
第二代是聊天。
第三代不仅仅是“更自主的编码”。
它是受控的委派。
你不仅会让 AI 编写代码。
你会让智能体调查错误、审查差异、检查文件、总结仓库状态、运行测试、比较选项、制定计划并并行处理隔离的子任务。
为了实现这一点，智能体需要内置的判断力和内置的工具。
而不是一个空壳，等着你手动螺栓式地连接每一个功能。
Mercury 仍然支持技能。
这一部分很重要。
当你需要专业化功能时，技能非常棒。
技能可以编码可重复的工作流、团队约定、特定工具的流程或特定领域的行为。
但技能应该是扩展智能体。
它们不应该用来弥补核心的空虚。
核心智能体应该已经懂得如何运作。
它应该知道如何阅读。
如何检查。
如何询问。
如何跟踪。
如何记忆。
如何停止。
如何在限制范围内运作。
Mercury 的“第二大脑”就是其中的一部分。
README 将其描述为使用 SQLite 和 FTS5 全文搜索的持久化结构化记忆，具有记忆类型、自动提取、冲突解决和合并功能。(GitHub)
这很重要，因为智能体不应永远是无状态的工人。
如果你不得不不断重复你的偏好、项目上下文、习惯、目标和约束，那就是系统在浪费你的时间。
智能体应该记住重要的事情。
但记忆也需要边界。
这就是为什么 Mercury 的记忆是本地化的，并且可以通过像 /memory 这样的命令进行检查。(mercury.cosmicstack.org)
同样的哲学也适用于令牌（Token）。
智能体系统会很快消耗令牌。
长上下文。
工具调用。
文件扫描。
重复规划。
后台任务。
子智能体。
如果智能体没有预算意识，用户就要为混乱买单。
Mercury 包含每日预算强制执行、超过阈值后的自动精简行为，以及用于检查、重置、设置或覆盖使用情况的 /budget 命令。(GitHub)
同样，这不是一个附带功能。
它是让智能体每天都能被使用的组成部分。
真正的智能体需要记忆。
真正的智能体需要工具。
真正的智能体需要权限。
真正的智能体需要进度可见性。
真正的智能体需要停止按钮。
真正的智能体需要预算控制。
真正的智能体需要在安装任何额外东西之前就能发挥作用。
这就是我们所说的“开箱即用（自带电池）”。
不是臃肿。
不是鲁莽。
不是封闭。
只是足够有能力从第一次运行开始就服务于它的目的。
然后，当你需要更多时，再扩展它。
这就是我所相信的模式。
内置核心能力。
通过技能实现专业化工作流。
用户控制无处不在。
Mercury 仍处于早期阶段，但这是我们正在构建的方向：
一个不畏惧实际工作的智能体。
一个不需要每个基本动作都有单独技能的智能体。
一个能记住重要事情的智能体。
一个在行动之前会询问的智能体。
一个可以从 CLI 或 Telegram 运行的智能体。
一个可以在后台工作、委派给子智能体，但仍然受控于你的规则的智能体。
这就是我想要的未来。
不是空洞的聊天机器人。
不是鲁莽的操作员。
不是插件的迷宫。
一个有边界的能力型智能体。
Mercury Agent 是开源的。
GitHub:
https://github.com/cosmicstack-labs/mercury-agent
## 相关链接
- [Zaid](https://x.com/Ctrl_Alt_Zaid)
- [@Ctrl_Alt_Zaid](https://x.com/Ctrl_Alt_Zaid)
- [18K](https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711/analytics)
- [mercury.cosmicstack.org](https://mercury.cosmicstack.org/docs/reference/built-in-tools)
- [GitHub](https://github.com/cosmicstack-labs/mercury-agent)
- [mercury.cosmicstack.org](https://mercury.cosmicstack.org/docs/cli-commands/in-chat-commands)
- [mercury.cosmicstack.org](https://mercury.cosmicstack.org/docs/cli-commands/in-chat-commands)
- [GitHub](https://github.com/cosmicstack-labs/mercury-agent)
- [mercury.cosmicstack.org](https://mercury.cosmicstack.org/docs/cli-commands/in-chat-commands)
- [GitHub](https://github.com/cosmicstack-labs/mercury-agent)
- [https://github.com/cosmicstack-labs/mercury-agent](https://github.com/cosmicstack-labs/mercury-agent)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [3:29 PM · May 16, 2026](https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711)
- [18.8K Views](https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711/analytics)
- [View quotes](https://x.com/Ctrl_Alt_Zaid/status/2055551179799863711/quotes)
---
*导出时间: 2026/5/17 10:19:28*