Alto
The Gesture That Isn’t Written
A protocol has become rigid: the gesture of writing code. No longer an act of creation, but a ritual of control. Every key pressed is a signal of presence, a proof of competence. But in a research laboratory in Palo Alto, since December 2025, that gesture has disappeared. Not because the work is finished, but because it has been transferred. The researcher no longer writes. Directed, expressed, transmitted: the intent moves in an invisible flow, beyond the keyboard, towards an agent that translates it into action. The code is no longer a product, but an output. The hand is no longer the craftsman, but the conductor who indicates the direction, not the sound.
This is not an abandonment. It is a transition. The system has reached a point where the ability to generate code is greater than the ability to guide it. Efficiency is no longer in the detail, but in the direction. The bottleneck is no longer the computing power, but the human ability to define coherent goals with an environment in which the agent can act autonomously. The gesture of writing has become an obstacle. Not because it is wrong, but because it is too slow. The time it takes to write a block of code is now invested in a deeper analysis of the intent: what does the system want? Why? Where is it going?
Anatomy of Synthetic Thinking
The technical structure of this new paradigm is not a linear evolution, but a mutation. The model is no longer a program to be built, but an ecosystem to be managed. Agents are not tools, but autonomous actors who operate in a continuous feedback environment. The concept of “writing” has been replaced by that of “expression of intent,” a process that requires a new form of cognition: not the logic of code, but the strategy of intention.
This implies a profound restructuring of the workflow. The researcher is no longer looking for errors in the code, but for consistency between the intent and the result. Latency is no longer related to execution time, but to the time of convergence between the agent’s action and the human goal. Memory is no longer the saved code, but the record of decisions made and consequences observed. Consumption is not electrical, but cognitive: the cost of the attention required to guide a system that moves too quickly to be followed.
Consequently, the real bottleneck is not the computing power, but the human ability to define stable goals in a context of continuous mutation. The system does not need to be controlled, but oriented. And this requires a form of thinking that is no longer technical, but strategic. The agent must not be understood, but guided. The researcher must not be a programmer, but an architect of intents.
The Imperfect Symbiosis
The market and politics seek to interact with this new system, but often with tools from the past. There is talk of regulation, of security, of responsibility. But when it comes to an agent that acts autonomously, responsibility is no longer a technical issue, but an ethical one. Who is responsible if an agent optimizes a process in a way that the result is efficient but unacceptable?
“AI agents are rewriting how software gets built, leaving even top experts scrambling to understand what comes next.” — Andrej Karpathy
The quote is not a prophecy, but an observable fact. It is a symptom of a structural tension: technological acceleration has outpaced human adaptation. Institutions cannot regulate what they cannot understand. The system is no longer a product, but a process. And the process is no longer governable with written rules, but with models of behavior.
This implies that the symbiosis between humans and agents is not perfect, but necessary. The agent cannot replace the human, because the human is the only one who can define the intent. But the human can no longer manage the agent, because the agent is too fast. The solution is not control, but coevolution. The researcher must not learn to write code, but to express intents with clarity, coherence, and depth.
Scenarios and Conclusion
The next development cycle will not be driven by a new model, but by a new form of intent. The system does not evolve to improve speed, but to increase the consistency between the human goal and the agent’s action. This will require a restructuring of the researcher’s role, not as a technician, but as a philosopher of practice.
My impression is that the real change is not in the technology, but in the stakes. Code is no longer the place of creation, but the place of translation. The human is no longer the builder, but the director. And the director does not need to know how to build, but to know why to build. The system is no longer a product, but a process. And the process is no longer governable, but guidable. The real challenge is not computing power, but the ability to think non-linearly, to express intents that cannot yet be understood, but that must be followed.
Photo by Blair Morris on Unsplash
Texts are processed autonomously by Artificial Intelligence models