The era of GPT agents

Agentic AI systems, designed to autonomously pursue complex goals with limited supervision, represent a turning point in the capabilities and applications of AI technologies. Unlike traditional systems that perform narrowly defined tasks, agentic AI systems are distinguished by their adaptability, reasoning, and ability to independently execute multi-step actions to achieve user-specified goals.

Characteristics of Agentic AI Systems

Agentic AI systems are defined by:

• Complexity of objectives : pursuit of sophisticated objectives.
• Environmental adaptability : taking into account dynamic environments.
• Autonomous execution : reliable performance with minimal supervision.

Examples : optimization of supply chains, personal assistants capable of planning meetings or managing communications without human intervention.

Associated risks

1. Unintentional errors : costly or non-compliant actions without prior approval.
2. Malicious exploitation : diversion for cyberattacks or disinformation.
3. Social inequalities : replacement of jobs, economic imbalances.
4. Systemic failures : dependence on similar algorithms, shared vulnerabilities.

Approaches to responsible integration

1. Test reliability : rigorous simulations to identify failure modes.
2. Constraint of actions : human approval for high-stakes actions.
3. Transparency : clear logs of decisions and reasoning.
4. Stop mechanisms : systems enabling secure interruption of operations.
5. Public-private collaboration : define common standards.

Increased productivity and efficiency

Agentic AI systems promise to automate repetitive and complex tasks, increasing productivity. According to McKinsey, AI could add up to $13 trillion to the global economy by 000, largely thanks to these systems.

Scalability
Unlike traditional systems, agentic AI systems can operate at an unprecedented scale. For example, in radiology, an agentic AI system can not only detect anomalies in medical images but also produce detailed diagnostic reports, reducing the workload of professionals.

Acceleration of scientific discoveries
These systems are already contributing to breakthroughs in areas such as drug discovery and climate modeling. DeepMind's AlphaFold, for example, has revolutionized the prediction of protein structures, a task previously considered extremely complex.

Operational challenges

Risks of unintentional actions
The autonomy of agentic AI systems can lead to unforeseen errors. For example, a scheduling assistant could mistakenly book expensive, non-refundable flights, highlighting the importance of control mechanisms.

Abuse and vulnerabilities
These systems can be exploited for malicious purposes, including cyberattacks or the dissemination of disinformation. This requires rigorous security measures to limit these risks.

Indirect impacts
The adoption of these systems could exacerbate social and economic inequalities. Jobs could be lost, and cybersecurity imbalances or systemic risks could emerge, particularly in the event of failures correlated with homogeneous algorithms.

Risk reduction: good practices

1. Assessing fit to context
   AI systems must undergo exhaustive testing under realistic conditions. For example, autonomous driving systems must prove their reliability in rare and high-stakes scenarios.
2. Limit autonomous actions
   Establishing human approval for critical actions, such as large financial transactions or legal commitments, helps reduce risk.
3. Minimize errors by default
   Systems should operate with reasonable default behaviors, avoiding costly decisions without explicit user consent. Techniques like Reinforcement Learning with Human Feedback (RLHF) can help align their behavior with user intentions.
4. Ensure transparency
   Systems must produce clear and understandable logs of their actions and reasoning. Recent advances in chain-of-thought facilitate this transparency, although challenges remain in terms of accuracy and complexity.
5. Set up automatic monitoring
   Secondary AI systems can monitor the behaviors of primary systems to detect anomalies or dangerous behaviors. However, this raises concerns about cost, scalability, and privacy.
6. Implement attribution mechanisms
   Assigning unique identifiers to AI agents can facilitate their traceability. For example, financial transactions made by an AI system could be linked to its human manager via these identifiers.
7. Provide stopping mechanisms
   Systems must be designed to allow for safe interruption of their operation if necessary. Pre-established contingency plans can ensure continuity in critical applications.

Societal and economic consequences

Changes in the workforce
Agentic AI systems could replace jobs in sectors once considered safe, including creative and managerial occupations. A PwC study estimates that nearly 30% of jobs could be automated by the mid-2030s. Training workers and sharing the benefits fairly will be essential.

Systemic failures
Overreliance on similar algorithms can lead to shared vulnerabilities, amplifying failures. Diversifying AI development methods and planning for robust fallback systems are essential measures to avoid these risks.

Economic transitions
The use of agentic AI systems in management and creative tasks could reshape the job market. Investing in education and training in new skills will be crucial to help individuals adapt.

Towards effective regulation

Regulatory needs
Legislative frameworks such as the European Union's AI Act, focused on risk-based regulation, offer a promising foundation. Policymakers must collaborate with businesses to establish clear and effective guidelines.

Encouraging diversity in AI
Relying on a single development approach increases risks. Promoting varied methods can make systems more robust and secure.

Public-private collaboration
Multi-stakeholder initiatives, such as the Partnership on AI, illustrate the importance of collaboration between industry, academia and policy makers to accelerate the adoption of security practices.

Continuous research and evaluation
Continued efforts to understand the technical, social, and economic implications of agentic AI are needed. Organizations like OpenAI and DeepMind play a key role in developing security protocols and evaluation measures.

Agentic AI systems offer unprecedented opportunities, but their widespread adoption will depend on our ability to balance innovation with responsibility. Through rigorous governance and diverse approaches, we can leverage their potential while minimizing risks.

To go further: how AI agents operate in autonomous systems, autonomous agents and global intelligence and our presentation of AI agents.