返回首页

对比阅读

对比阅读:More Tools, Worse Agents: Selection, Not Quantity, Is the Key 与 Agent 接入上百个工具仍不好用,关键在筛选而非数量

AEN
MCPSkill HubQwen3-4B·

More Tools, Worse Agents: Selection, Not Quantity, Is the Key

An Agent connected to hundreds of tools may still fail to improve its task completion rate. The problem may not be the model, but how many tools it “sees” in each turn. The original article reminds us that more tools do not necessarily mean better usability: what determines success is the selection and presentation strategy.

What this is

An Agent is an AI assistant that can call tools step by step to accomplish a goal. It typically works with a set of tool descriptions covering functionality, required fields, and usage formats. Showing all descriptions at once consumes the model’s limited information-processing capacity; tokens can be roughly understood as the unit of textual information a model can process.

Two things need to be distinguished here. The capability model determines whether a tool is built as a specialized function, a general-purpose command, or a Skill—a set of instructions explaining how to perform a particular type of task, optionally with attached scripts. The presentation strategy determines how many tools the Agent sees at the current moment. A more common design is to provide a short directory first, expand the relevant descriptions as the task progresses, and even let the Agent search for tools on demand when needed.

MCP (Model Context Protocol, a standard interface for connecting AI applications to external services) solves the problem of how tools connect. It does not decide which tools should be exposed in a given turn. Skill Hub is closer to a distribution channel for operating instructions.

Industry view

We believe the industry is shifting from “connection count” to “tool governance.” For builders, the key metric is not how many services they have connected, but whether they can find the right tool, clearly say “not found” when no suitable option exists, and restrict the data and system permissions available to tools.

On-demand discovery, however, has not been fully validated. The test cited in the original article used Qwen3-4B and covered only three tasks. Both the all-at-once and on-demand discovery groups completed all three tasks, achieving an accuracy rate of 100%. The test shows only that active discovery reduced the volume of exposed descriptions and the measured execution time, while also producing irrelevant calls and premature termination. It cannot support the claim that on-demand approaches generally improve accuracy.

The risks are equally concrete: third-party tool descriptions may contain prompt-injection instructions, and scripts attached to Skills may execute local code. Searchability and installability do not mean that access to contacts, email, or other permissions should be granted immediately. Sources, versions, and access scopes should all be reviewed.

Impact on regular people

For enterprise IT: When evaluating AI assistants, procurement teams should shift their focus from tool count to selection mechanisms, invocation logs, and permission boundaries, avoiding products that offer “many features but a chaotic selection process.”

For individual professionals: An Agent purchased by an employer is not necessarily better because it connects to more tools. A more reliable assistant understands what the current task requires and can stop and explain why when no suitable tool is available.

For the consumer market: Connections between applications will become more widespread, but the difference users actually experience will not be “how many services are supported.” It will be whether an assistant requires fewer steps, completes tasks accurately, and gives users control over sensitive-data permissions.

来源: juejin.cn
BZH
MCPSkill HubQwen3-4B·

Agent 接入上百个工具仍不好用,关键在筛选而非数量

一个 Agent 接了上百个工具,任务完成率却没有提高,问题可能不在模型,而在它每轮“看见”了多少工具。原文提醒我们:工具多不等于更好用,决定成败的是选择与展示策略。

这是什么

Agent 是能分步骤调用工具完成目标的 AI 助手。它通常会面对一套工具说明,包括功能、必填字段和使用格式。一次性展示全部说明,会占用模型一次能处理的信息空间;token 可粗略理解为模型处理文本的信息篇幅单位。

这里需要区分两件事:能力形态决定工具做成专用功能、通用命令还是 Skill(写明某类任务如何操作、也可附带脚本的说明文件);展示策略决定当前给 Agent 看多少工具。更常见的设计是先提供简短目录,任务推进后再展开相关说明,甚至由 Agent 在需要时主动查找工具。

MCP(Model Context Protocol,连接 AI 应用与外部服务的通用接口)解决的是工具“怎样接得上”,不负责判断这一轮该展示哪些工具。Skill Hub 则更像操作说明的分发渠道。

行业怎么看

我们认为,行业正在从“连接数量”转向“工具治理”。对搭建者而言,更重要的指标不是接入多少服务,而是能否找到正确工具、遇到不匹配时是否明确说“没找到”,以及是否限制工具可访问的数据和系统权限。

但按需发现并非已被充分证明。原文引用的测试使用 Qwen3-4B,仅覆盖三道任务,一次性展示与按需查找两组都完成了 3/3,准确率均为 100%。测试只显示主动发现减少了说明暴露量和实测用时,同时出现无关调用与过早结束,不能据此断言按需方案普遍提高准确率。

风险也很具体:第三方工具说明可能夹带诱导性指令,Skill 附带脚本还可能执行本地代码。可搜索、可安装不等于可以立即开放通讯录、邮件或其他权限,来源、版本和访问范围都应被检查。

对普通人的影响

对企业 IT:评估 AI 助手时,采购重点应从工具数量转向筛选机制、调用记录与权限边界,避免“功能很多、选择很乱”。

对个人职场:企业采购的 Agent 未必接入工具越多越好。更可靠的助手通常知道当前任务需要什么,也能在缺少合适工具时停止并说明原因。

对消费市场:未来应用间连接会更普遍,但用户真正感知到的差异将不是“支持多少服务”,而是能否少点几步、准确完成任务并控制敏感数据授权。

来源: juejin.cn