Once the number of tools exceeds roughly 100, even advanced models are more likely to select the wrong tool. The bottleneck to deploying Agents is shifting from “whether it can do the job” to “whether it can reliably invoke the right capability.”
What this is
Here, tools refer to the programmatic interfaces an Agent can call—the “hands and feet” through which AI interacts with files, web pages, email, command lines, and other systems. The article divides tools into five categories: perception tools that read information, execution tools that modify external systems, tools that collaborate with people and other Agents, communication tools that proactively message users, and event-triggered tools activated by external events such as new emails or system alerts.
In the past, developers often built a dedicated tool for every capability, such as separate tools for reading PDFs, Word documents, and PowerPoint presentations. The article argues that once tools exceed roughly 100, the model’s selection burden and error rate rise sharply. One alternative is a “Skill + general-purpose executor”: use natural-language documentation to spell out the steps, then have a code interpreter or terminal execute them. Capabilities with complex parameters and stable rules remain better suited to dedicated tools; tasks with frequently changing details and clear procedures are better expressed as Skills.
MCP—the protocol that lets AI applications connect to external tools and data sources in a standardized way—solves the interface compatibility problem. Skills solve the problem of how capabilities are organized and selected. The two are not interchangeable.
Industry view
We see tool governance becoming a new dividing line in Agent productization. When there are few tools, the underlying model determines the task ceiling. When there are many, interface design, permission controls, error recovery, and testing determine whether the system runs reliably. The broader the task coverage of general-purpose Agents such as Manus and OpenClaw, the larger their tool catalogs become—and the more prominent this contradiction becomes.
But “reducing tools and replacing dedicated interfaces with a general-purpose executor” is not cost-free. Natural-language Skills are generally less stable than structured functions, and complex parameters are more likely to be passed incorrectly. During high-risk operations, a general-purpose executor may expand the scope for mistakes. Execution tools can send email, modify files, or run commands; if the model chooses the wrong step, the business damage may far exceed the cost of one inaccurate answer. Dedicated tools should therefore remain in high-risk areas such as payments, deletions, and publishing, with risks controlled through explicit parameters, permission boundaries, and human confirmation.
Another risk is that permission governance may lag behind capability expansion as Agents gain more tools. Putting file reading and production-system writes in the same execution environment may seem convenient, but it gives a single prompt error or tool mis-selection greater impact. Whether a tool is useful depends not only on how many there are, but also on whether permissions can be isolated by task, invocation scopes can be restricted, and audit records can be retained.
Impact on regular people
For enterprise IT: As enterprises deploy Agents, the focus is shifting from purchasing models to mapping tool inventories, data permissions, and operational logs. Especially once email, cloud storage, customer service, and financial systems are connected, interface governance and human-confirmation mechanisms deserve more investment than polished demonstrations.
For individual professionals: Personal assistants will gradually move from answering questions to executing work across applications—for example, researching information, generating a document, and submitting it to a colleague. In the short term, people will still need to handle permission requests, review results, and establish responsibility. Tasks should not bypass review simply because an Agent executed them.
For the consumer market: Assistants that can proactively send messages, run on schedules, and collaborate across channels will be more likely to become part of regular usage habits. But incorrect messages, repeated actions, and faulty operations will also become more common. When choosing such products, consumers should look beyond task-completion rates and examine whether confirmation mechanisms, data-access scopes, and operation records are clear.