MAS offers the chance for AI to operate in harmony, orchestrated with precision and speed to deliver beautiful results, say Patrick Meyer, artificial intelligence senior architect and technical director at Sopra Steria UK, and Clément Bénesse, senior AI researcher at opsci.ai.
Picture a symphony orchestra. Each musician is a specialist: the violinist has mastered their strings, the pianist knows every key, the conductor orchestrates the whole performance. Now imagine asking a single musician to play every instrument simultaneously. The result would be chaos, not harmony.
This analogy captures the fundamental shift happening in artificial intelligence today. While the tech world has been racing to build ever-larger, more powerful monolithic models, a different approach is quietly gaining ground: Multi-Agent Systems (MAS). Instead of relying on a single, all-purpose AI model to handle every task, MAS deploy small, specialised agents that collaborate to achieve faster, more precise results.
This is a paradigm shift so could the future of AI might be more collaborative than we imagined?
Breaking down the monolith: How MAS differ from traditional LLMs
From a user's perspective, the interface remains familiar. "For a user, nothing changes at the interface level, we're still dealing with the same types of dialogue interactions," explains Patrick Meyer, artificial intelligence senior architect and technical director at Sopra Steria UK. "What changes is the system's ability to respond to different types of questions."
The fundamental difference lies beneath the surface. "Instead of having a generative model that directly produces output from a prompt (what we might call a static or monolithic model), we have work happening in the background, querying multiple models multiple times, allowing the system to build higher-quality output," he continues.
However, Clément Bénesse, senior AI researcher at opsci.ai, nuances this approach. "While in theory, you could allow every agent to talk with everyone without any structure, in practice you want to limit prolonged interactions and enforce some kind of information flow. This naturally leads you to a more structured architecture."
This shift unlocks new possibilities for complexity and precision. "We're moving from a system where you ask simple questions like 'summarise this text' to much more complex requests like 'create a PDF summary of last quarter's sales compared to the competition,' and you get a complete report in just a few minutes," Meyer says.
Behind the magic: How multi-agent systems work
"The architecture of a multi-agent system is relatively stable," he explains. "We have an orchestrator that coordinates the sequence of agents, understanding user intent and dispatching tasks to specialised agents."
This creates a new paradigm where agents communicate in natural language rather than technical APIs. "Unlike microservices that call each other via APIs in technical language, everything here is exchanged in natural language," Meyer notes.
The technical foundation often relies on what experts call Directed Acyclic Graphs (DAGs). Bénesse explains the concept: "Think of it as a workflow where each task has clear dependencies. You can't start analysing sales data until you've collected it, but you can search for competitor information in parallel. This structure prevents infinite loops while maximising efficiency."
Take a complex request like generating a quarterly sales report. "The system breaks this into a clear sequence," Bénesse continues. "First, an agent interprets the request. Then, two agents work in parallel. One searches the internet for competitor data while another queries internal databases. Once both finish, a third agent analyses and compares the data, a fourth generates the PDF, and a final agent validates the output."
This orchestration creates unprecedented efficiency. "The key advantage is identifying the 'critical path': the longest sequence of dependent tasks that determines minimum execution time," Bénesse adds. "Good parallelisation can dramatically reduce overall response time."
Efficiency through specialisation
One of the most compelling advantages of MAS lies in resource efficiency. "We can use much smaller models, what we call SLMs (Small Language Models). These models have about 10-12 billion parameters and generally run on a small graphics card consuming a few dozen watts," Meyer explains.
This represents a huge advantage compared to resource-hungry giants. While monolithic models require H100 GPUs to operate, specialised agents can run on consumer-grade hardware. "It's a kind of 'divide and conquer': we have small models that interact with each other, with guardrails that check if the produced content is acceptable," he says.
Bénesse highlights additional advantages: "Beyond lowering costs, MAS enable natural hybridisation between LLMs and existing tools. The era of isolated LLMs is over, we can now integrate all the legacy scripts and systems companies have developed over years."
The biological analogy is striking. "It's like the human brain, we don't activate it 100% all the time, and not in the same locations for the same tasks," Meyer observes.
However, this efficiency comes with trade-offs. The increased interactions between agents generate consumption. "The question is whether one inference on a large model equals ten inferences on small models,” says Bénesse. “We can also allocate more resources to complex tasks or stop calculations early if intermediate results aren't promising. This adaptive resource allocation is key."
Addressing the elephant in the room: Do MAS reduce hallucinations?
The relationship between multi-agent architectures and AI reliability is nuanced. "It's a complex question," Meyer admits. "RAG can reduce hallucinations by providing the right information, but the simple fact that agents interact with each other won't necessarily improve this problem."
The intuitive assumption that more agents equal better results doesn't always hold. "It's not because you put two people in a room that you get twice as much intelligence," Meyer notes.
Yet MAS introduce capabilities that monolithic models lack. "Multi-agent systems introduce an interactivity that monolithic models don't have. They can ask questions to clarify or refine," Meyer explains.
Bénesse offers a different perspective on reliability: "We can implement cross-validation between agents. One agent's output gets examined by another specialised in quality control. When multiple agents can approach the same task, consensus mechanisms help aggregate proposals for more robust results."
He also emphasises MAS's ability to access real-time information: "One of LLMs' strengths is transforming unstructured material into structured data. Agents can now access the internet directly, thus reducing hallucinations from outdated knowledge."
The security challenge
As these systems grow more complex, new challenges emerge. "The difficulty for industry adoption is making agents robust and secure," Bénesse warns. "More agents mean a bigger attack surface, and we're still learning how stable these pipelines are in production environments."
The transparency of agent interactions offers some advantages for monitoring, but traditional auditing approaches fall short when dealing with multiple interconnected models. Developing robust LLM auditing frameworks capable of tracing decision paths across agent networks has become critical for enterprise adoption.
This represents one of the key hurdles for widespread adoption: ensuring that collaborative AI systems remain as reliable and secure as their monolithic predecessors.
The collaborative future of AI
"The revolution is already in motion," Meyer observes. This shift toward specialised, collaborative AI represents more than technical evolution: it's a path toward democratised AI development. Smaller models are easier and cheaper to train, making advanced AI capabilities accessible beyond tech giants.
"We'll see the emergence of business agents (HR agents, procurement agents, product management agents) that will talk to each other to produce interesting results," Meyer predicts.
Bénesse concludes with measured optimism: "While challenges remain around security and stability, the potential for more efficient, transparent, and accessible AI is compelling.”
The future of AI isn't about building bigger orchestras: it's about teaching musicians to play together more beautifully.