AI Agents: Towards a New OS for Secure Orchestration

The Necessary Evolution of Operating Systems for AI Agents

AI agents, increasingly present in the professional world, require a new approach to operating systems. Unlike traditional software, these agents do not merely execute predefined functions. They interpret intentions, plan actions, request tools, and autonomously interact with information systems. This growing complexity calls for the creation of a new "OS" that goes beyond managing hardware resources, orchestrating the intentions and actions of AI agents.

For decades, enterprise computing has been structured around a simple principle: applications execute functions, users trigger actions, and operating systems manage resources. However, the arrival of AI agents disrupts this logic. An agent does not simply respond to a question. It can interpret an intention, plan multiple steps, call tools, query data, produce a deliverable, solicit human input, correct its trajectory, and sometimes act directly within the information system. In other words, the AI agent is no longer just a conversational interface. It becomes a form of software actor.

Thus, a central question arises: do we need a new OS for AI agents? The short answer would be: yes, but not in the traditional sense. We probably do not need a new Windows, Linux, or macOS for agents. We need a new layer of execution, governance, and orchestration capable of making agents reliable, observable, secure, and scalable. An Agent OS, in short. Not a machine operating system, but an operating system of intention, context, actions, and trust.

The Challenges Posed by AI Agents

Traditional software follows a deterministic logic, producing predictable results from specific inputs. In contrast, AI agents operate on probabilistic principles, integrating dynamic contexts and varied objectives. They require supervision of decisions rather than applications, and a management of rights that extends to entities capable of interpreting intentions. This implies traceability of reasoning and intermediate decisions, rendering traditional software management approaches obsolete.

Traditional software is relatively predictable. It receives an input, executes deterministic logic, and produces an output. Even when complex, its behavior relies on written, tested, versioned, and audited code. An AI agent functions differently. It combines probabilistic reasoning, dynamic context, tool calls, memory, permissions, objectives, constraints, and sometimes a form of operational autonomy. This changes everything.

In a traditional system, applications are supervised. In an agent-based system, decisions must be supervised. In a traditional system, rights are granted to users or services. In an agent-based system, rights must be granted to entities capable of interpreting an intention and acting across multiple systems. In a traditional system, API calls are traced. In an agent-based system, chains of reasoning, used sources, called tools, human validations, context errors, and intermediate decisions must be traced.

Towards an "Agent OS": Initiatives and Developments

Several major players in the tech sector are converging towards the idea of an "Agent OS," although the term is not always explicitly used. Microsoft, for example, offers Azure AI Foundry and its Agent Service to design, deploy, orchestrate, and supervise AI agents, integrating connections to tools like SharePoint, Microsoft Fabric, Azure AI Search, and action connectors via Azure Logic Apps. OpenAI is developing an Agents SDK for the orchestration and validation of agents, distinguishing between approaches where the LLM dynamically decides the flow and those where orchestration remains code-controlled. Anthropic promotes the Model Context Protocol to standardize integrations between agents and external tools, aiming to reduce integration fragmentation and avoid the need to create a specific connector for each model-tool-data combination. Additionally, the A2A protocol, initiated by Google and now under the auspices of the Linux Foundation, enables secure communication between agents from different frameworks.

The Essential Functions of an Agent OS

A true "Agent OS" should fulfill several key functions to ensure effective integration of AI agents within businesses. It must manage the identity of agents, which could be numerous and specialized, such as legal or commercial agents. Permission management is crucial, as agents do not merely converse; they act, necessitating a policy engine to frame their actions. Orchestrating multiple agents is also essential, as is managing memory and context, which must be balanced to avoid drift. An Agent OS should enable governance of this agent memory.

Finally, supervising reasoning and securing the execution of tools are imperatives to ensure the reliability and security of AI agents. Agent observability includes monitoring the documents used, tools called, and intermediate decisions. Security is likely at the heart of the matter. An agent connected to tools becomes a potential risk vector if adequate security measures are not implemented.