In the race toward automation, sophistication has become a trap
Every week brings its share of spectacular demonstrations. Autonomous agents capable of booking a trip, managing a project, analyzing documents, or coordinating multiple systems. Companies are stepping up their experiments, and consultants are outdoing one another in creativity to design ever-more-complex architectures.
However, a subtle trend is beginning to emerge among organizations that are actually deploying these systems on a large scale: the more sophisticated the agents become, the less predictable they are.
This paradox is worth examining closely.
For years, software development was guided by a relatively simple logic: adding features generally enhanced a product’s capabilities. Generative artificial intelligence works differently. Every new tool, every new data source, and every new rule added to an agent also increases its uncertainty space.
In other words, power increases along with vulnerability.
When Less Is More
One of the most counterintuitive lessons observed in recent months in the AI ecosystem concerns the deliberate reduction of complexity.
Some technology companies have begun removing a significant portion of the tools made available to their employees—not to limit their capabilities, but to improve their reliability.
It seems counterintuitive. Why remove features from a system that’s supposed to handle more tasks?
Because an agent doesn’t always choose the optimal tool. The broader its repertoire of actions becomes, the more it must weigh different options. Every additional decision creates an opportunity for error.
The issue, therefore, is no longer the model’s power. The latest generations of large language models already have capabilities far exceeding those utilized by most companies.
The real challenge comes down to orchestration.
The luxury industry is already familiar with this approach
This tension is no stranger to luxury brands.
A fine watchmaking workshop does not endlessly add complications simply to demonstrate its expertise. An additional complication adds value only when it serves a specific purpose.
The same logic now applies to artificial intelligence systems.
The mistake many organizations make is to view the platform as a collection of possibilities. Every new connector, every new database, and every new feature seems to represent progress.
In practice, these additions often create more noise than value.
The best agent isn’t necessarily the one who can do everything. It’s the one that perfectly accomplishes what it was designed to do.
The End of the Fantasy of the Universal Agent
Since the emergence of LLMs, one vision has captured the public’s imagination: that of an assistant capable of replacing a multitude of human tasks thanks to its virtually unlimited autonomy.
This promise is driving a large portion of current investments.
But feedback from the most advanced companies tells a different story.
The systems that generate the most value today are rarely the most ambitious ones. They are generally specialized agents that operate within a clearly defined scope, with a limited number of possible actions and measurable success criteria.
Sophistication then shifts elsewhere.
Not in the number of tools available, but in the quality of the evaluations.
Competitive advantage no longer lies in the model
For two years, the industry focused primarily on identifying which model performed best.
GPT, Claude, Gemini, Qwen, and Mistral were the main topics of conversation.
This issue is gradually becoming less important.
Models are now advancing at such a rapid pace that improvements in their capabilities can sometimes destabilize systems built on their previous behaviors.
Competitive advantage is therefore shifting to a different arena: the ability to measure, test, and manage results.
The most mature organizations invest less in increasing the number of agents and more in creating protocols capable of evaluating their quality.
The true strategic infrastructure of AI may not be the agent itself.
The system checks to make sure it’s working.
The Elite of the Selection
Artificial intelligence ultimately confronts us with an age-old question.
When everything becomes possible, what should you choose not to do?
For a long time, the digital world has emphasized abundance: more data, more content, and more features.
AI takes this logic to the extreme.
Faced with this proliferation of possibilities, a new form of excellence is emerging: the ability to eliminate.
Eliminate unnecessary tools. Eliminate poorly defined tasks. Eliminate workflows whose value no one can clearly explain.
In the world of luxury, exclusivity often stems from selection.
The future of AI agents could follow the same pattern.
The organizations that will get the most out of this technology may not be the ones that build the most agents.
But those who know which ones not to build.
Cette publication est également disponible en :
