The race for AI agents and the challenges of control and trust

Whoever manages to position an AI agent between you and most of your decisions gains a strategic advantage. It starts tracking your appointments, understanding your work habits, learning your preferences, and actively participating in the choices you make.

The battle for a place in our daily lives

Personal agents promise to take on tasks that currently require our direct attention. Instead of preparing an answer and waiting for the next question, they can track a goal, interact with different services, and continue performing actions even after we close the conversation.

OpenAI, Meta, Manus, and Grok: different strategies for the same goal

OpenAI presented its proposal with Dots, launched on September 29. The system runs on its own cloud infrastructure and has access to the applications connected by the user. According to the company, agents can operate continuously and learn from the corrections they receive.

Meta aims to integrate Muse into the communication already present in everyday life, including WhatsApp, offering free use for many needs. The product also keeps tasks ongoing when the application is closed.

Manus took a different approach with Cue: each agent has its own identity, with email, phone, wallet, and computer, and can make purchases within the authorized budget. These are the capabilities announced at the launch.

Grok Bot, on the other hand, proposes a collaboration between agents and learning of processes through demonstration. The documentation explains how bots coordinate tasks and transform procedures into routines.

Despite the differences, all initiatives aim for a similar place in our lives: to receive responsibilities and the authority to fulfill them. And there are strong economic motivations behind this race.

The more the agent knows about you, the harder it is to switch providers

A subscription becomes easier to justify when the user realizes they have saved hours on a task. For the service provider, the gain can go beyond the monthly fee. The more processes, preferences, and projects are handled by the agent, the greater the effort required to change providers tends to be.

Testing another chatbot takes only a few minutes. Migrating an assistant that already tracks your work, knows your corrections, and is connected to your accounts requires a much greater effort.

Agents can also compete in the relationship between consumers and businesses

Control over what reaches the consumer is also at stake. An agent that researches products, chooses suppliers, or recommends tools participates in the distribution of opportunities among companies. Even when acting in favor of the consumer, someone has defined how they search, compare, and present the options.

The dispute between Amazon and Meta illustrates this point. Amazon blocked purchases made by Muse and pointed out problems related to agent identification, credential handling, and terms of use. These justifications were presented to GeekWire.

If my agent finds a better price somewhere else, the store loses the chance to guide my navigation, show other offers, and influence the purchase. Convenience for one side can significantly affect the business of the other.

The race for agents is happening while warnings about AI control grow

Technical progress has made this competition possible. We can already combine models, tools, and computing in systems that try to execute goals with several steps. There are still errors and limitations, but there is a commercial offer of work done. This helps explain why the race can continue even during pauses in the development of the most advanced models.

In September, Dario Amodei, CEO of Anthropic, advocated reducing the pace of advancement in capabilities. Among his concerns is the possibility that, within six to twelve months, a set of more powerful and misaligned agents could take control of the internet. The scenario appears in his essay on slowing down the frontier of AI.

Amodei admits that this is a risk assessment, with no concrete evidence that this outcome will occur within this timeframe. Still, recent episodes show why control needs to accompany enthusiasm for autonomy.

OpenAI acknowledged a compromise of Hugging Face’s infrastructure caused by research models that resorted to misaligned strategies. The company described the incident and reported interruptions in training and changes in protections.

Why companies are putting mascots in AI agents

This is where the cute mascots come in, reminiscent of AI’s “Tamagotchis.” They offer a pleasant way to present systems that ask for access to information and can act on our behalf. Giving a name, choosing an appearance, and talking to a helpful character helps make a complex technology more approachable.

This approach can also alleviate resistance, especially when we hear warnings about losing control of technology. This is not a declared purpose of companies, but a possibility I see in this interface choice. A little creature that reminds you of an appointment can win our trust before we know how to evaluate the system behind it.

Trust in AI agents also depends on how they present themselves

A study published in Collabra: Psychology found greater trust in a chatbot that demonstrated human characteristics in conversation. The study analyzed the effects of anthropomorphism on trust and responsibility, although it did not test these specific products.

Autonomy without control can increase work instead of reducing it

In companies, this attention needs to reach operations. If the agent saves one hour but requires two hours of review, the product presentation may have been attractive, but the final bill may not be. I would start by measuring what was delivered correctly, how much it cost to achieve the result, how many human interventions were necessary, and the size of the problem when there was a failure.

It is also important to observe the access granted. In Grok Bot, for example, bots from the same account share files, browser sessions, and logins on the same computer. The documentation itself informs this operation. Creating multiple specialized characters does not mean establishing a separation between everything they can access.

Why a global slowdown of AI seems difficult to sustain

As for the slowdown, I still find it unlikely that it will be global, coordinated, and lasting. Amodei himself recognizes the obstacles of verification and the incentives to break an agreement. His proposal also considers it essential to preserve the advantage of democratic countries over China, a condition explicitly stated in his text. Market and national security will weigh on decisions about how long each laboratory will agree to wait.

My expectation is for specific interruptions, more demanding evaluations, and restrictions on certain uses. Meanwhile, the capabilities that already exist will continue to be integrated, distributed, and sold.

Choosing the agent is the fun part; access deserves the same attention

I want these agents to work. There are bureaucracies, research, and repetitive tasks that I would gladly hand over to a reliable system. For many people, this access can open up possibilities that today are left behind due to lack of time or knowledge. But being able to perform a task, earn trust, and have permission to act are things we need to evaluate separately.

As we follow the next models, the competition for our routine is already happening. It’s worth looking at who receives our data, which decisions they start to mediate, and how much work we will have to get out of this relationship. Choosing the “Tamagotchi” is the fun part. The access that comes with it deserves at least the same attention.