Wrapping every API—the interface through which software requests services from other software—as a tool has become a common approach to AI Agent development. But this article puts forward a more important judgment: what determines whether an Agent can complete real work is not how many interfaces it connects to, but whether those tools map to task objectives, permission boundaries, and execution outcomes.

We are seeing AI Agents move from “providing recommendations” to “acting on our behalf.” They can find files, read content, modify documents, and deliver results. But to cross that threshold, models need more than comprehension: they need tools for search, reading, editing, sending, collaboration, and scheduled triggers.

What this is

Simply put, an Agent is an AI assistant that can work toward a goal in steps. The model handles understanding and judgment, while tools provide the specific actions it is permitted to take. The article divides tools into five categories: searching emails and feedback forms falls under sensing information; writing a report is an execution action; calling other assistants or requesting human review enables collaboration; proactively sending a report is user communication; and starting automatically based on a schedule or new message is event triggering.

Design methods are changing as well. Developers are no longer starting only from API endpoints. Instead, they are designing interaction entry points around goals, an approach known as ACI (Agent-Computer Interface). For example, unifying PDF, Word, and PPT reading into a single “read document” tool is more reliable than asking an Agent to assemble several granular interfaces on its own.

Every tool must also specify when it should be used, what it must not do, how its parameters should be filled in, and what limitations apply to its results. Permissions for reading and sending must be separated in particular: viewing ten emails has limited impact, whereas sending ten emails creates ten real actions in the outside world.

Industry view

We believe the article’s most valuable contribution is elevating tool design from an engineering detail to a product and security issue. A general-purpose command line paired with a Skill—a set of operating instructions for an Agent—works well for tasks with variable workflows. A specialized tool with a fixed structure is better suited to payments, complex forms, and granular authorization. The former is flexible; the latter is easier to audit.

The risks are equally clear: an Agent may choose the wrong tool, misunderstand input, or quietly alter it. If a system automatically replaces quotation marks in a document, the Agent may struggle even to identify why it failed. Another objection is that wrapping tools around goals increases development and maintenance costs. Enterprises with numerous interfaces and substantial platform differences may not be able to organize enough stable, standardized entry points quickly.

Impact on regular people

For enterprise IT: Tool design will directly affect deployment costs and incident rates. Organizations need to assess task fit, permission isolation, and operational traceability—not just model accuracy.

For workplace professionals: Future AI assistants may do more than answer questions; they may research materials, revise documents, and send notifications on our behalf. An employee’s most valuable role will shift from executing every step toward defining objectives, validating results, and handling exceptions.

For the consumer market: Products that can genuinely complete bookings, expense reimbursements, or document organization will enter everyday use sooner. But payment, sending, and similar actions must retain confirmation mechanisms. As convenience increases, responsibility boundaries for erroneous actions will attract greater scrutiny.