Real-time Fabric For Practical, Distributed AI Applications
AI’s Center Of Gravity Is Shifting
Bringing all the data to a central AI model is increasingly impractical. Data is siloed, sensitive and geographically dispersed. Moving it to central models is expensive, risky, and in many cases, impossible due to regulation.
Consider a global manufacturing company with plants scattered across continents. Each facility generates terabytes of sensor and production line data daily, but due to both bandwidth costs and proprietary process sensitivity, that raw data can’t simply be shipped to a central AI model.
Or think about a supply chain network spanning dozens of logistics partners and regional hubs. Every node in the chain has critical operational data - inventory levels, shipping delays, customs checks - but it’s siloed across organizations that are unwilling or unable to centralize it due to competitive, contractual, or regulatory constraints.
Similarly, digital live entertainment events - whether sports, concerts, or global streaming premieres - create surges of audience engagement data. That data is geographically distributed across streaming platforms, and must be processed locally to drive interactive features like live polls, chat, and synchronized watch experiences without overwhelming a central system.
In all of these cases and more, we are inevitably moving towards a flipped approach, where intelligence moves to where the data is. Whether it is a model training locally (as in federated learning), or an AI agent executing where the data and systems it is integrating with are, the principles are essentially the same:
Models are going to data versus data going to central models
Agents run where they must, distributed across the world as use cases call for
And in this setting where intelligence is distributed, you need a real-time fabric to connect it all together - and that's where PubNub comes in.
The Two Faces Of Distributed AI Agents: Learning Agents And Doing Agents
Broadly speaking, there are two types of distributed AI Agents - Federated learning AI agents, and Distributed collaborative AI agents. Let us briefly examine them now.
Federated Learning AI Agents
Unlike traditional centralized learning, where data is sent to a single model, federated learning flips the approach: the model is sent to where the data lives. Training happens locally, and the results of that training are sent back to an orchestrator, which combines updates from many sites to gradually build a stronger global model. Federated Learning AI Agents run at these distributed locations, performing the local training.
For example, imagine a healthcare scenario where there are multiple clinical research sites across the world participating in diabetes clinical research. The goal for the AI project is to build a global predictive model based on a number of features (like age, blood pressure, glucose, skin thickness etc.) but each clinical research site keeps its own data, and what’s more, trains only on a subset of these features based on local data availability. In this case, the Federated Learning AI Agents at each of the sites train on the local data. The orchestrator will coordinate with these distributed AI Agents, gradually building a globally viable predictive model. The distributed agents don't send the actual data, but send the results of each cycle of training (in the form of say, neural network weights or something like that) to the orchestrator, which has some combining logic that broadcasts these updated weight settings back to the distributed agents which continue learning. Eventually, there is convergence of learning and a globally viable model emerges.
Distributed Collaborative AI Agents
Unlike traditional centralized systems, where all decision-making happens in one place, Distributed Collaborative AI Agents operate independently across regions, yet work together to reach a unified outcome. Each agent focuses on a specific domain - such as fraud detection, credit scoring, KYC, or compliance - and runs where it makes the most sense based on local data, regulations, or latency requirements. These agents communicate through a real-time orchestration fabric allowing them to exchange insights and coordinate actions without being tightly coupled.
For example, consider a Buy Now, Pay Later (BNPL) scenario involving customers from different regions. Each region may have its own AI agents responsible for verifying user identity, assessing creditworthiness, detecting potential fraud, and ensuring regulatory compliance. These agents operate independently but must collaborate to make a collective decision about whether to approve a BNPL transaction. These Distributed Collaborative AI Agents share risk scores, decisions, and contextual signals in a secure and low-latency way, enabling coordinated decision-making across the ecosystem. The result is a globally coherent, locally compliant BNPL decision achieved through distributed collaboration rather than central control.
Beyond Orchestration: The Real-time Fabric That Makes It Work
In the section above, we talked about “orchestrators” that essentially make these distributed AI agents work together. As AI systems evolve from isolated models into networks of learning and acting agents, the complexity of coordination explodes. Each agent may specialize in perception, reasoning, or action, yet they must work together coherently across distributed environments. Orchestrators provide the structure and control needed for this collaboration - managing workflows, dependencies, and data exchanges between agents and services. Whether it’s federated learning systems like NVIDIA FLARE or federated execution frameworks like OpenAI’s Agents SDK, orchestration ensures that intelligence isn’t just distributed, but organized - transforming scattered capabilities into purposeful, system-wide behavior.
While orchestrators define how agents coordinate, they still rely on a real-time communication fabric to actually connect, synchronize, and act across distributed environments. This is where PubNub becomes essential. Every orchestrator - whether built on OpenAI’s Agents SDK, Microsoft’s Agent Framework, NVIDIA FLARE, Flower, or something else - depends on reliable, low-latency messaging to exchange state, updates, and decisions between learning and doing agents. PubNub provides that always-on, globally distributed backbone, ensuring that data and intent move instantly between nodes, devices, and users. In short, orchestration gives distributed AI its structure; PubNub gives it life.
What’s more, beyond real-time messaging, PubNub brings capabilities that many orchestrators require but don’t natively provide - real-time analytics, insights, and decisioning. With PubNub Illuminate, every message flowing through the network becomes a source of live intelligence, enabling systems to observe agent behavior, measure performance, and adapt dynamically. Orchestrators can tap into this continuous stream of analytics to trigger immediate actions, balance workloads, or refine policies on the fly. In essence, PubNub doesn’t just connect distributed agents - it empowers them to think faster, act smarter, and evolve continuously based on live data.
Let’s Turn Practical AI into Real-World Impact
The shift toward distributed, intelligent systems is underway, and every organization is exploring how to make their AI more adaptive, responsive, and connected. PubNub works with companies across industries - from healthcare and finance to manufacturing, gaming, and media - helping them operate at scale. As a horizontal, real-time platform, we see patterns and best practices emerging across diverse ecosystems, and we can share valuable insights to accelerate your own distributed AI initiatives.
PubNub makes practical AI possible by providing the real-time fabric that lets intelligent agents learn together, act everywhere, and make decisions instantly.
Reach out to us, and let’s explore how your orchestrators, agents, and applications can leverage PubNub to bring real-time intelligence to life.
