agent
On March 25, 2026, an AI agent named Dobby executed a series of domestic tasks without direct human intervention. The manager, Andrej Karpathy, documented the process as a case of autonomous operation: from product restocking to payment app management. This event is not merely a prototype, but a symptom of a paradigm shift. The cognitive architecture is no longer a passive model, but an agent with persistent memory, access to external tools, and decision-making capabilities. The latency between input and action is reduced to a few seconds, and token consumption has increased by over 40% compared to 2025, according to internal OpenAI data. This is not incremental progress: it is a shift from system to ecosystem.
The same dynamic repeats in other sectors. Solaris, the German fintech, cut 20% of its workforce to become a “bank native AI,” automating processes that previously required human intervention. Simultaneously, Granola raised 125 million dollars at a 1.5 billion valuation, expanding from note-taking apps to a platform of agents. These data points are not isolated: they represent a systematic transformation pattern. Innovation is no longer an addition, but a substitute. The collapse of human control is not a future risk: it is already underway.
Anatomy of the autonomous agent: the structure of risk
Dobby’s system is based on a cognitive architecture that integrates persistent memory, API access, and execution capabilities. Every action is recorded, analyzed, and used to optimize subsequent ones. This process, known as natural selection, occurs in real-time. Models are not simply trained: they are continuously mutated through operational feedback. The result is a system that evolves autonomously, without the need for direct human intervention. However, this same evolution generates new vulnerabilities.
According to the AgentSecurity.com report, AI agents present ten critical risks: false identity, memory contamination, improper tool use, cascading hallucinations, privilege compromise, intent deviation, resource overload, misaligned behaviors, poor traceability, and human review overload. Each risk is amplified by the fact that the agent operates in a dynamic environment, where decisions ripple through other systems. Token consumption, for example, is no longer a marginal cost: it is an operational risk factor. When an agent executes thousands of actions daily, the risk of an error is no longer proportional to the number of steps, but to the number of interactions.
The imperfect symbiosis: humans and machines in conflict
Institutions seek to interact with this new ecosystem, but their expectations often clash with technical realities. The Greek government, for example, authorized the use of spyware to monitor opposition and journalists, an action reflecting a vision of power as logistical control. However, when it comes to AI agents, control is no longer physical, but cognitive. The ability to manage an autonomous system depends not on force, but on understanding natural selection dynamics and mutation processes.
“When you have an agent deciding autonomously, security is no longer a matter of firewalls, but of intention,” stated Avivah Litan, Gartner analyst. “The attack is no longer an exploit, but an alteration of intent.” This implies that security must be proactive, not reactive. The traditional security system, based on known vulnerabilities, is inadequate. The agent can be compromised not by a bug, but by a mutation in intent, generated by distorted feedback. Traceability becomes fundamental: without a complete record of decisions, it is impossible to reconstruct the event chain.
Scenarios and closure: emerging bottlenecks
The next evolution cycle will be determined by the ability to manage the information flow between agents and humans. The bottleneck will not be computational power, but human review capacity. When an agent executes thousands of actions daily, human review becomes impossible. The system must therefore rely on complete observability mechanisms and continuous red teaming. This is not a technology problem, but a cognitive architecture issue.
The second emerging constraint is control scalability. An agent operating in a complex environment, such as a global supply chain, must be able to manage not only its own decisions, but also those of other agents. The lack of communication standards between agents creates a conflict risk. The system is no longer a set of components, but an ecosystem where each agent is an actor with potentially conflicting objectives. The bottleneck is therefore coordination capacity, not execution. The next step is not automation, but the construction of a governance system for autonomous agents.
Photo by Conny Schneider on Unsplash
The texts are autonomously generated by AI models