“The emergent is unlike its components insofar as these are incommensurable, and it cannot be reduced to their sum or their difference.” - G.H.Lewes, 1875 (Lewes coined the term “Emergent”).

The Past: From Emergence to Orchestration

Emergence is a compelling phenomenon observable both in natural systems and in engineered designs, where complex behaviors and patterns arise from simple interactions. As James McClelland and John Holland have illustrated, systems comprised of simple agents can evolve to exhibit intricate patterns and capabilities that transcend those of any individual component.

Murray Gell-Mann, Nobel Laureate and co-founder of SantaFe Institute studied complex adaptive systems and delved into how emergence plays a role in everything from particle physics to biological systems. He believed that understanding the emergent properties of complex systems could provide insights into how the universe works at both the smallest and largest scales.

^{Figure 1.}

Imagine millions of birds following just three simple rules: avoid crowding your neighbors (separation), fly in the same direction as those around you (alignment), and stay close to the flock (cohesion). Nothing in any individual bird’s “programming” calls for the complex murmuration pictured above. Yet, when many starlings form a system, their repeated basic interactions give rise to this breathtaking behavior [Figure 1]. This is emergence in action – the birth of complex phenomena from surprisingly simple, non-linear rules.

This principle extends far beyond the elegance of birds in flight. Consider the seemingly random motion of gas particles. Governed by simple kinetic theory, their interactions create the macroscopic phenomena you rely on – pressure and temperature. Even tiny variations in initial conditions, like the infamous "butterfly effect" in atmospheric convection, can cause vastly different weather patterns, highlighting the non-linear nature of these emergent systems.

The Reaction-Diffusion model, particularly exemplified by the Belousov-Zhabotinsky (BZ) reaction, is a classic example of how simple local chemical reactions and diffusion lead to oscillations in chemical concentrations, and the oscillations propagate as waves through the medium and as these waves interact, they form complex spatial patterns.

In The Society of Mind, Minsky describes agents as the fundamental components of the mind, each performing specific tasks. He conceptualizes the mind as a society of these small components or agents, where each agent is responsible for different mental functions. These agents work both independently and cooperatively, leading to the emergence of intelligent behavior and cognitive processes through their interactions.

The power of emergence isn't limited to the physical world. Cellular automata, originally conceptualized by Stanislaw Ulam and Jon Von Neumann in 1940s, with basic birth/death rules for individual cells (exemplified by Conway's Game of Life [Figure 2]), exhibit complex patterns, mimicking real-world phenomena. Similarly, population dynamics, with basic predator-prey interactions, results in boom-and-bust cycles. Ant colony optimization algorithms model [Figure-2] the pheromone-trail-update and probabilistic decision-making process of individual ants, leading to the emergence of efficient solutions to complex optimization problems further demonstrating how simple rules can govern intricate spatial, temporal, and decision-making patterns.

$ \textrm{NextState}(x,y)\left\{\begin{matrix}1 & \textrm{if}(x,y)\textrm{ has exactly 3 live neighbors} \\1 & \textrm{if}(x,y)\textrm{ is alive and has 2 or 3 live neighbors} \\0 & \textrm{otherwise} \\\end{matrix}\right. $

This kind of swarm intelligence relies heavily on efficient communication and information exchange between the agents for orchestrating actions and forming coherent global patterns. For decades intelligent software agents have relied on structured message passing, standardized protocols, ontologies, middleware solutions, coordination languages, and both direct and indirect communication methods.

The evolution of programming languages has been marked by a quest to achieve higher levels of abstraction, enhancing accessibility for humans and fostering system interoperability. This journey began with the adoption of declarative, functional, structured, and object-oriented programming. In the 1980s, Donald Knuth introduced the concept of literate programming, proposing that programs should be viewed and written as literary works. The 1990s saw another advancement with Yoav Shoham's agent-oriented programming, which integrated goals and beliefs directly into program structures.

More recently, LLMs have further transformed this landscape by significantly reducing the barriers for agents to communicate with each other and with humans in natural language. This (a) reduces complexity by dramatically simplifying the integration architecture; (b) increases scalability as adding a new system will no longer require updating every other system and (c) enhances flexibility as systems can be updated, replaced, or reconfigured with minimal impact on the overall network of integrations.

The Present: Why should we care now?

We are entering the Agentic Era, marked by an explosive proliferation of intelligent agents globally. The rapid growth of LLM-based agents is already evident. To manage these distributed intelligences effectively in the near future, it is crucial to recognize and learn from recurring historical patterns. Historically, computing has continuously evolved from monolithic systems to complex distributed architectures, necessitating advanced orchestration and routing. This was important for scalability, which in turn meant interoperability, composability, reusability and most importantly discoverability.

^{Figure 3.}

As shown in [Figure-3], in the personal computing era, the shift from centralized mainframes to distributed computing was managed through innovations like TCP/IP routing. This evolution deepened in the Internet era, as static websites and monolithic servers transitioned to dynamic server clusters and microservices, requiring sophisticated service routing techniques. The cloud computing era brought further decentralization with distributed databases, enhancing data management through advanced slicing and transaction routing. Today, in the Agentic Era, we see the need to orchestrate LLM-based distributed agents. Orchestration of LLM-based agents is crucial for enhancing task specialization, continuous learning, resource optimization, and collaborative problem-solving. By directing tasks to specialized agents best suited for them, orchestration improves both efficiency and output quality, leveraging each agent's strengths and compensating for weaknesses. This setup not only fosters synergy among agents but also enables adaptive learning from interactions, refining strategies over time. Moreover, orchestration dynamically adjusts task routing and resource allocation based on real-time feedback, maintaining high performance amid changing conditions. It also facilitates the decomposition of complex problems into manageable components, promoting innovative solutions through collaborative efforts. Inherent feedback loops in orchestration refine system behaviors and foster the development of new operational dynamics, enhancing the system’s robustness and resilience.

Orchestration in the agentic era will unlock unprecedented problem-solving capabilities.

Enterprise agents are already beginning to revolutionize the IT delivery across various sectors by unlocking system-level intelligence and driving significant productivity gains. In customer service, agents integrated with CRM systems like Zendesk and Salesforce efficiently manage inquiries and support. Marketing agents enhance engagement and conversions by generating personalized content across multiple platforms. For financial services, agents offer strategic investment advice, perform market analyses and invoice processing. HR assistants streamline recruitment by automating candidate screening and initial interviews. In legal sectors, agents expedite document review and compliance monitoring, while in healthcare, LLM-based agents are positively impacting patient care by enhancing diagnostic accuracy and personalizing treatment plans, thereby improving outcomes and operational productivity. Supply chain coordinators leverage these agents to predict demands, manage inventory, and handle logistical challenges. Additionally, security agents within the enterprise agent ecosystem can ensure data privacy and system integrity throughout these processes. These applications demonstrate how enterprise agents are crucial in enhancing IT delivery, ensuring more efficient and effective business operations across industries.

The Future: Emergent Enterprise Systems

Building upon our foundational approach to generative AI, we envision a future where every enterprise system is integrated within intelligent agents. This proliferation of agents has the potential to unleash emergent system-level intelligence, thereby unlocking extraordinary problem-solving capabilities. But to achieve this, we must address the following critical challenges:

Discoverability: In a rapidly evolving agent ecosystem, the ability to efficiently locate the appropriate resources or services becomes crucial. The challenge of discoverability among intelligent agents hinges on the ability to efficiently locate and evaluate the right agent for a specific task amidst a vast and diverse ecosystem. This complexity is compounded by the dynamic nature of agents and underlying models (LLMs, LVMs), whose capabilities and availability may evolve in real-time. Discoverability must address issues like matching detailed task requirements with agent functionalities, managing interdependencies among multiple agents, and overcoming inconsistencies in metadata and lack of standardization. Similar to how the Domain Name Service (DNS) translates human-readable domain names into IP addresses to facilitate efficient coordination across the internet, we need robust mechanisms that enable the discovery of agents based on various constraints such as cost, accuracy, performance, and safety.

Interoperability (Inter-agent Communication): For deeply intelligent systems to emerge, different agents must seamlessly communicate and collaborate. LLMs significantly lower the barrier to inter-agent communication, facilitating smooth data exchanges and integrating human insights into decision-making processes. This not only enhances operational speed and quality but also fosters a collaborative ecosystem where complex problem-solving is expedited and the synergy between humans and AI is maximized.

Continuous Learning: It is essential for agents to continuously evolve through learning from interactions and feedback. By establishing robust feedback loops, agents can receive and act upon inputs from users and fellow agents. Inspired by Edsger Dijkstra’s principles of self-stabilizing systems and structured programming, our approach ensures that agents can autonomously correct and refine their behaviors, which is vital for maintaining fault tolerance in distributed systems.

Observability (Behavioral Monitoring): Continuous monitoring of agent behaviors and interactions is key to detecting emergent patterns and phenomena. Advanced analytics and visualization tools will need to employed to scrutinize the vast data generated by agents, helping to identify and leverage emergent behaviors and intelligence.

Orchestration: As the complexity of tasks increases, the need for sophisticated orchestrator systems that can manage and coordinate multiple agents becomes more critical. Drawing inspiration from Butler Lampson’s work on network architecture, conceptualized decades ago, we aim to integrate principles of network switching and system-level routing into orchestrating LLMs. However, our approach goes beyond just routing to the best models or agents. Our orchestrator system is designed to effectively manage and coordinate entire workflows within enterprise systems. This enhances automation and efficiency, improves decision-making, and provides scalability, driving significant value to enterprise operations. Our orchestrator system is poised to lead this transformation, ensuring that agents operate in sync to elevate collective intelligence and optimize operational efficiency across enterprise workflows

The advent of LLMs has significantly altered the trajectory of enterprise AI. With the introduction of agent orchestration, we are poised at the brink of another major evolutionary leap, one that promises to foster truly emergent system intelligence.

At Emergence, we are uniquely positioned with deep system-development and AI expertise and real-world experience to operate across three critical layers of research, development and deployment as shown in Figure 4. At the foundational level, our research focuses on advancing the science of Large Language Models (LLMs) in areas such as reasoning, planning, self-improvement, and long-term memory. The middle layer is dedicated to customer-specific model development, such as fine-tuning LLMs for specialized domains and tasks. Here, we build sophisticated agents tailored to specific enterprise needs. The top layer involves orchestrating these agents—both our homegrown ones and others from industry sources as appropriate. Our first scaled offering, the agent orchestrator, serves as a dynamic system to manage and optimize the interaction of these diverse agents, ensuring efficient task execution and seamless integration across various applications and systems. This comprehensive approach allows us to drive innovation and deliver tailored AI solutions that address complex challenges in the ever-evolving landscape of generative AI.

^{Figure 4.}

Conclusion

In the evolving landscape of technology, Emergence stands at the forefront, harnessing the power of LLMs through sophisticated orchestrators that intelligently route queries to the most effective systems or agents. As we’ve seen, this approach draws inspiration from a rich tapestry of concepts including emergence, complex adaptive systems, distributed computing, agent-oriented programming, deep learning, programming abstractions and design patterns.

At the core of our development is a commitment to social intelligence—the idea that our technologies should enhance the capabilities of human teams and networks. By integrating diverse LLMs and computational agents, we are crafting a distributed, intelligent framework that is more than just a tool; it is a partner in the quest for knowledge and efficiency.

In this journey, we are not just engineers or developers; we are architects of cognitive ecosystems, crafting tools that think, learn, and ultimately, understand. Join us as we explore the frontier of artificial intelligence, driven by the legacy of giants and the promise of tomorrow's innovations.