Déploiement de l'IA agentique devançant les cadres de gouvernance

AI systems are increasingly deployed as autonomous agents — executing multi-step tasks, browsing the web, writing and running code, making purchases, interacting with APIs, and operating computer interfaces — with minimal human oversight between steps. This represents a qualitative shift from AI as a tool that responds to individual prompts to AI as an actor that pursues goals across extended action sequences, where errors compound and unintended consequences accumulate.

The deployment of agentic AI accelerated rapidly in 2025-2026. Anthropic released Claude computer use capabilities (October 2024), enabling AI to operate computer interfaces autonomously. OpenAI launched Operator (January 2025), an agent that performs web-based tasks on behalf of users. Google DeepMind deployed Mariner for web browsing and Jules for coding. Coding agents achieved dramatic capability gains: on the SWE-bench benchmark (resolving real GitHub issues), performance rose from under 5% in early 2024 to over 50% by mid-2025, with the best systems now resolving issues that would take experienced developers hours. Companies like Cognition (Devin), Factory, and others raised hundreds of millions of dollars for autonomous coding products.

The International AI Safety Report 2026 explicitly identifies agentic AI as an emerging risk category, noting that "the deployment of AI systems as autonomous agents introduces novel risk vectors including compounding errors across action sequences, difficulty of attributing responsibility for agent actions, and potential for agents to take actions with irreversible real-world consequences." The report notes that safety evaluation methodologies developed for single-turn interactions are inadequate for agentic systems that operate over extended time horizons.

The risk structure is distinct from other AI hazards. In a standard AI deployment, the human reviews each output before acting on it. In agentic deployment, the AI takes a sequence of actions — searching, reading, clicking, typing, submitting — with the human seeing only the final result. Each step has some probability of error or misinterpretation; across a sequence of dozens or hundreds of steps, errors compound. An agent that misunderstands a task specification may take confident, well-executed actions in the wrong direction — purchasing the wrong items, sending incorrect communications, modifying the wrong files, or interacting with the wrong systems — before the human can intervene.

Accountability gaps are structural. When an AI agent sends an email, makes a purchase, modifies a database, or files a form on behalf of a user or organization, who bears responsibility if the action is incorrect, harmful, or unauthorized? Existing legal frameworks assume a human decision-maker at each point of action. Agentic AI disrupts this assumption without providing an alternative accountability structure.

Multi-agent dynamics add complexity. As organizations deploy multiple AI agents that interact with each other — one agent's output becoming another's input — emergent behaviours can arise that no individual agent was designed to produce. Market dynamics, information cascades, and coordination failures become possible at machine speed without human intervention points.