At a European distribution centre, a critical incident occurs. Deliveries are at risk of being delayed. Within seconds, around 10 agents reorganise the entire logistics chain: reallocating stock, renegotiating delivery slots, readjusting priorities. The twist? No human intervention was required: the agents in question are software powered by artificial intelligence. This scenario, already operational at several major industrial companies, illustrates the power of agentic AI. But it also raises a legitimate concern: what happens if these agents make decisions with irreversible consequences? How can we ensure that a system capable of self-organisation remains under control? Michel Poujol, AI Expert Lead and CTO of the rAIse programme at Sopra Steria, examines the orchestration mechanisms that prevent autonomy from degenerating into chaos.
A technological breakthrough?
Multi-agent systems are not new. Since the 1990s, researchers and industrialists have explored architectures where multiple software entities collaborate to solve complex problems. But today's agentic AI represents far more than an incremental evolution of these pioneering systems. "What distinguishes current multi-agent systems powered by generative AI from those of the 1990s is less their ability to scale industrially, than their capacity to now perform highly complex tasks that were previously reserved for humans," explains Poujol.
This breakthrough is based on the reasoning capabilities of language models embedded within agents. Unlike traditional expert systems that follow predefined rules, these agents can analyse novel situations, adapt their strategy and even question their own decisions. "They act somewhat like a human brain that can reconsider its previous decisions if things don't go as planned," he notes. However, this parallel with human intelligence has its limits: "This brain is generally much less powerful in terms of adaptability than a human brain, and safeguards are often necessary."
So what exactly are these agents for? "Agentic AI considerably extends the possibilities for using AI in general and generative AI in particular, to use cases that we couldn't have imagined just a few months ago," emphasises Poujol. Tasks that require several weeks of work from specialised teams can now be accomplished in a few hours, without writing a line of code, simply by controlling these systems in natural language.
The orchestrator at the heart
At the centre of any agentic system sits an orchestrator that plays a crucial conductor role. In most current applications, this orchestrator uses a language model to define tasks to be accomplished through reasoning, then selects the appropriate tools to execute them. "Each task is then specified in a prompt detailing its context. The selection of the tool to perform this task is made from the verbal definition provided in its docstring – typically a few sentences describing the tool, located in its code header," details Poujol.
Fortunately, there are multiple solutions to make it more reliable and robust. For example, by using rule-based AI or hybrid approaches that combine generative AI with symbolic AI: systems based on explicit logic rules and structured knowledge representations, rather than statistical learning. Where generative models reason through pattern recognition in training data, symbolic AI operates through formal reasoning on predefined concepts and relationships. This combination leverages the flexibility of language models while adding the reliability of deterministic logic.
Failure management constitutes another major challenge. When an agent fails in its task or returns an inconsistent response, the system must be able to react appropriately. This is where the concept of Operational Design Domain (ODD) comes in, which is gaining importance in AI-based systems engineering. "It's a notion that will likely become increasingly important in the engineering of AI-based systems or applications," confirms Poujol.
An ODD precisely defines the functional scope of each system component. "If these scopes are well defined, any behaviour outside this scope by an agent or AI model must lead the application to execute another component on another specifically defined scope for handling this type of situation," he explains. This approach enables the management of hallucinations and erratic behaviours, which can therefore be avoided for most AI applications through good design. Nevertheless, there remain AI applications, such as those for general use, where it is impossible to fully circumscribe their functional scope and where consequently these problems are much more difficult to avoid.
Between autonomy and control: a precarious balance
The question remains of where to place human control in a system where autonomous agents reason in parallel and exchange information. For Poujol, the answer must again be provided during the specification and design phases.
Agent autonomy must never be synonymous with lack of control: "We can and must put in any system, whether AI-based or not, as many human interactions as is necessary to meet its specifications or regulatory requirements," insists Poujol. "Just because it's now possible to integrate models that can reason or make decisions in place of humans into these systems doesn't mean we're obliged to give them complete freedom."
So how do we deploy this type of system harmoniously? Here are some points to consider within the confiance.ai programme framework – an approach in which Sopra Steria participates, ensuring that human supervision questions are systematically addressed and resolved in a structured manner.
Controlled deployment
To deploy agentic systems responsibly in critical sectors, several safeguards are essential. "One of the main causes of these failures comes from poor specification of use case requirements," says Poujol. "Without this adoption by business teams responsible for expressing requirements, it is practically impossible to succeed in an AI project."
The degree of autonomy granted to each agent must be clearly defined in the specifications. The scope of admissible decisions must be precisely circumscribed, and any decision outside this scope must lead to its transfer to another agent or a human.
Agentic AI promises spectacular productivity gains and opens unprecedented prospects for automating complex tasks, but this promise can only be fulfilled on condition of adopting rigorous engineering methodologies, integrating the necessary safeguards from the design phase. "In principle, no serious system should be deployed with chaotic behaviour above an admissible threshold that is generally very low," adds Poujol. In production systems, good orchestration conquers chaos. Between autonomy and control, orchestration does not underline mastery: it requires it. "It will be our responsibility to ensure it is for the better and not for the worse," concludes the expert.