The Smart Agent: From Just an Answering Robot to a Secretary that Books Flights and Hotels

After years of users of artificial intelligence robots getting used to asking questions and receiving answers, the fast-paced tech race today has shifted to a different model, where the intelligent system does not wait for a question but rather receives a "task" to execute.

This shift coincided with the emergence of what is known as "AI Agents," systems that can understand the user's goal, break it down into steps, utilize various tools, websites, and services, and continue working until reaching the desired outcome.

Instead of asking a chat robot how to organize dozens of files, one can assign the agent to organize them on its own, and instead of requesting steps to book a trip, it can search for flights and hotels, compare them, and then carry out the booking process.

Programming and professional work were among the most prominent environments where this shift appeared early on, as OpenAI launched its software agent Codex to perform software engineering tasks, while Anthropic expanded the idea beyond programming through Claude Cowork, which can handle various files and tools and execute complete work tasks, rather than just providing guidance to the user.

The emergence of the Cowork agent itself was the result of an interesting observation; Anthropic states that non-technical users began using Claude Code, originally designed for developers, to organize folders, remove duplicate files, and write spreadsheet formulas, prompting it to transfer the same capabilities to a more user-friendly interface.

A sample that included 1.2 million usage sessions for Cowork during last May showed that 33.4% of them were related to business operations, such as compiling scattered updates into a single report, preparing lists for new employees, and reconciling data within spreadsheets, compared to only 8.7% for software development.

But the next phase began to shift the agent from the work environment to life itself.

Here emerged the concept of Personal AI Agent, which does not require its owner to be a programmer or a professional, but connects to the email, calendar, messages, accounts, and services that the person uses, and then executes tasks on their behalf.

One of the prominent companies pushing in this direction is Instinct, founded by young researcher Noah Shin, who participated as a college student in the early Reflexion research on making AI agents learn from their mistakes, before leaving university, working at Sierra, a company specializing in agents, and then founding Instinct.

Despite relying solely on invitations, Instinct quickly spread among Silicon Valley investors, raising $250 million in a funding round that valued the company at about $2.5 billion, after its users began sharing their experiences on social media.

The reason was not that it was a language model providing better answers, but rather what the agent could actually do; it has been used by testers to book trips and restaurants, cancel subscriptions, purchase necessities, manage email messages, and organize calendars.

One user mentioned that they used the agent to review their subscriptions and negotiate them, find an important message in the spam folder, and complete missing appointments in their calendar.

In September, the company Instinct Concierge added capabilities for the agent to go beyond the computer screen to phone calls; it can call a restaurant that does not accept online reservations, try to get a last-minute canceled medical appointment, or communicate with a service company to resolve an issue with the user's bill.

The success of the smart agent idea has put the small startup in direct competition with 'Meta', which launched its personal agent Muse on September 8, in one of the most prominent attempts to transition agents from products used by specialists to a widely consumed product.

Meta summarizes the difference between the two generations by stating that Muse does not just engage in conversation, but can connect to the services the user relies on, such as email, calendar, payments, shopping, and restaurants, and then operate through them on their behalf.

The user can ask their smart agent to send an email, book a trip, attempt to lower a bill, fill out a form, convert a cooking video into a shopping list, or purchase a product and pay for it. If there is no API available for a needed service, the agent can handle it through a web browser.

As organizations expand their use of artificial intelligence agents, calls are increasing to establish controls that define responsibility for their actions, as well as the systems and data they can access.

In the same week that Instinct added calls, Meta also began allowing Muse to connect with American companies by phone, quickly shifting the competition from 'who has the smartest model?' to who can execute the largest part of users' daily tasks?

In the same race, Google launched a smart agent named CC, initially as an experiment to explore what it can autonomously perform within a real digital environment.

The company's vision was that the agent should not only answer questions or suggest steps but should have its own account, tools, and operating environment, enabling it to understand the user's context, access authorized services, and then execute tasks through them while keeping sensitive decisions under human control.

From this perspective, CC was designed as an experimental agent capable of working across various Google services, not as a separate chatbot; the idea since its launch was to test the transition of artificial intelligence from an assistant waiting for instructions to a system that can follow tasks, connect scattered information, and take the necessary steps to achieve a complete goal, while building a clear model of permissions, consent, and trust.

Google has taken this idea a step further; on September 17, the company expanded its experimental project CC from a personal agent to an agent for families and homes, which can be used by up to 6 individuals, each of whom has the ability to specify the information that the agent is allowed to access.

CC has its own Google account, connects to services like Gmail, Calendar, Chat, and Docs, operates on a separate cloud computer, and can gather school messages, activity schedules, bills, and household tasks into a shared daily summary, add appointments to the calendar, prepare shopping lists and meal plans, and even fill out activity registration forms after user approval.

Thus, the concept of the agent expands once again: from an agent that assists the programmer, to an agent that helps the employee, then an agent that manages part of a person's life, and finally an agent that an entire family relies on.

However, each step that makes the agent more useful also makes it need broader permissions; booking trips requires user data and possibly payment information, organizing mail requires reading messages, and managing the family may mean accessing children's schedules, school, and home.

The problem indeed appeared in the early trials of Instinct; users reported issues that included executing unintended actions, and the agent systems remain vulnerable to "Prompt Injection" attacks, which can attempt to deceive the agent through hidden instructions within websites or messages it reads.

For this reason, "Meta" built several layers of protection within Muse, separating sensitive login data from the environment in which the agent operates, and preventing the mail connector from passing verification codes, password reset links, and quick login links to it, along with dedicated layers to resist malicious instructions.

As for Instinct, it offers mechanisms to protect user data privacy by minimizing the exposure of credentials to the agent, restricting access to accounts and services to the necessary permissions to perform the task, while keeping the user in control of granting permissions and reviewing sensitive actions. Both companies acknowledge that agents, like any AI system, can make mistakes.

Here may lie the biggest shift in the AI race; after years of competition among companies over who has the chatbot that provides the best answer, the competition is gradually shifting to who has the "agent" that can accomplish the most in the user's digital life.