As more people focus on agentic AI, it’s becoming clear that there’s a big difference between experimenting and actually making things work in the real world. Many companies can show agents working together in test settings, but far fewer can run large, secure and reliable multi-agent systems in real businesses. On a recent episode of Futurise, Soo Ling Lim, CTO at Futuria, talked about what it really takes to move from single agents to what she calls multi-multi-agent teams.
Soo Ling’s views come from years of experience with agent-based modelling, long before large language models became popular. While LLMs have made agents much more capable in the last year or so by adding reasoning, role consistency and smoother workflows, they are just one part of a bigger system. Real agentic AI also needs strong orchestration, solid enterprise infrastructure and disciplined operations.
From Agents to Organised Teams
At its core, agentic AI is about autonomy, teamwork, and good oversight, not just having an agent present. Single agents or simple workflows might be interesting, but they don’t meet enterprise needs. Multi-agent teams need to delegate, coordinate, escalate and share information reliably, so they can achieve more than any single agent or model could on its own.
As these systems grow, they become much more complex. Soo Ling points out that when many agents interact, even small changes in prompts, models, or how things are organized can cause big shifts in behaviour. Figuring out, testing and checking these interactions turns into a systems challenge, not just a modelling issue.
Secure by Design, Not Added Later
In regulated and high-trust settings, security must come first. Agentic systems often handle sensitive data and connect to important business systems, so building security in from the start is essential. Soo Ling shares key ideas like zero-trust agent setups, detailed permissions, careful control over tools and data, and strong rules for how agents communicate with each other.
Observability is just as important. Agentic systems for enterprises need to be easy to audit, predictable when needed and fully traceable. Keeping logs of what agents see, think and do is key for compliance, fixing problems and building trust. User feedback also helps people stay in control, so they can step in, make corrections, and guide agent behaviour as needed.
The Next Frontier: Orchestration, Testing, and Optimisation
Looking forward, Soo Ling sees three big challenges for agentic AI. First is managing orchestration at scale, which means building flexible teams of agents that work together reliably. Second is evaluation, since the industry still needs strong testing tools to check not just single agents, but whole teams working in real situations. Third is optimization, which goes beyond just tweaking prompts to improving how agents work and delegate together.
As companies start to see agents as ongoing delivery units instead of just chatbots, their expectations change. Reliability, accountability and good oversight become just as important as what the agents can do. Multi-agent teams offer a strong new way to operate, but only if they are designed with the same care as any other important business system.
In this next phase, success will rely more on solid engineering and less on just trying things out, turning agentic AI from potential into something businesses can truly depend on.
