Industry executives and experts share their predictions for 2026. Read them in this 18th annual VMblog.com series exclusive.
By Neil Lawrence, Co-Founder & Chief Scientist, Trent AI and Professor of Machine Learning at the University of Cambridge
For the last several years, progress in large language models (LLMs) has been defined by scale. Bigger models, trained on more data with more compute, have reliably produced better benchmarks, smoother language, and broader capabilities. But that era is approaching a natural plateau. Large language models are already “good enough” for a very many real-world applications. The next major wave of innovation will not come from making models larger, but from learning how to orchestrate them more effectively.
This shift represents a fundamental change in how value is created with AI. Instead of asking, “How powerful is the model?”, the more important question becomes, “How well is the system designed around it?”
Models Are Commoditizing, Systems Are Not
Today, organizations have access to an unprecedented range of powerful foundation models. Differences still exist between them, but those differences are increasingly marginal compared to the gains that can be achieved through better system design. In practice, the choice between one strong model and another often matters less than how that model is prompted, monitored, combined with others, and embedded into a workflow.
This is where orchestration comes in. Orchestration is not simply about chaining prompts together. It is about designing frameworks that coordinate multiple agents, tools, and processes into a coherent system that reliably achieves a goal. These frameworks are becoming more accessible, which is both a strength and a risk. Much like the early days of home computing, the tools are now powerful enough that individuals can build sophisticated systems on their own. But without discipline, those systems can quickly become fragile, opaque, and hard to scale.
Orchestration as a New Skill Set
One of the most important implications of this shift is that orchestration itself becomes a core skill. Prompting a single model is already non-trivial, but orchestrating multiple agents introduces a new level of complexity. Developers must think in terms of roles, responsibilities, and interactions rather than isolated queries.
When done well, the results are extremely powerful. A multi-agent system allows for specialization: different agents can be assigned to planning, execution, verification, safety checks, or domain-specific reasoning. Crucially, these agents are interchangeable. If one agent underperforms in a given role, it can be swapped out for another, much like a coach changing players to better fit a strategy. This flexibility is possible because high-quality models are readily available across platforms.
What we are already seeing is that no single model is universally best. Performance varies depending on the task, the context, and even the way instructions are phrased. Orchestration lets practitioners exploit this diversity rather than being constrained by it.
From Single-Agent Tools to Agent Teams
Frameworks such as LangChain and similar orchestration layers are making it easier to build these systems. While the idea of “multi-agent AI” can sound futuristic or complex, in practice many useful systems do not require agents to negotiate or collaborate across organizations. Even within a single user’s environment, like on a laptop or internal server, agent orchestration can deliver dramatic gains in reliability and customization.
This accessibility means that innovation will often spread organically. Instead of being driven solely by large vendors offering monolithic solutions, orchestration techniques will propagate by word of mouth and shared experimentation. Individuals and small teams will discover that by managing their own “teams” of agents, they can outperform more rigid, one-size-fits-all products.
Indeed, their greatest advantage often comes from not relying on a single provider. Systems that orchestrate agents from multiple vendors are inherently more robust and adaptable. They reduce lock-in and allow organizations to evolve their AI stack as capabilities and costs change.
Language, Behavior, and Control
Another subtle but important dimension of orchestration is language itself. While English dominates most AI interactions today, there is no reason to assume it is always optimal. Different languages may elicit different behaviors or reasoning patterns from models. In an orchestrated system, it is entirely plausible that certain agents perform better when prompted in Mandarin, French, or another language. Orchestration frameworks make this kind of experimentation feasible.
More broadly, orchestration brings AI development within reach of individuals who are willing to “hack around” and experiment. The barrier to entry is no longer access to massive compute resources, but the ability to design systems thoughtfully. This democratization mirrors earlier technological shifts, where the real breakthroughs came not from hardware improvements alone, but from how people learned to use that hardware creatively.
The Decade of Agents
It is tempting to frame a specific year as “the year of agents,” but this misses the larger picture. Agent-based systems are not a short-term trend; they represent a long-term transformation. Even if model progress were to slow dramatically, or stop entirely, the potential of orchestration would remain. Better coordination, standards, safety mechanisms, and reliability tooling can unlock enormous value from models that already exist.
Looking ahead, new challenges will emerge. As organizations build multiple orchestrated systems, those systems will need to work together. The future may involve not just isolated agent teams, but networks of agents operating across departments, companies, and platforms, an “Internet or Intranet of Agents?”
Conclusion
The future of LLM innovation is not about chasing ever-larger models. It’s about mastering orchestration: designing systems that combine models intelligently, manage risk, and adapt to changing needs. In that future, competitive advantage will belong not to those with the biggest models, but to those who build the best teams, of humans and agents alike.
##
ABOUT THE AUTHOR
Neil Lawrence is Co-Founder and Chief Scientist of Trent AI. Neil has over 25 years of experience with machine learning models and the real-world applications of this technology. Prior to his role at Trent AI, Neil was the director of machine learning at Amazon. Neil is also the inaugural DeepMind Professor of Machine Learning at the University of Cambridge, where he continues to work on research at the crux of machine learning and systems, and is a visiting professor at the University of Sheffield. As well as his departmental roles, he serves on the board of the conference AISTATS and the ELLIS Foundation, and as the founding and series editor for the Proceedings of Machine Learning Research.
