Why AI agents need an event-driven architecture

By 2029, IDC projects that the number of AI agents deployed globally will reach 1.15 billion, more than 40 times the number of agents deployed in 2025.¹ As enterprises bet their future on agentic AI, the risk of not getting it right becomes the stuff of nightmares for business leaders. And the risk is real, especially if enterprises treat agentic AI as just another technology.

AI agents, by nature of their autonomous decisioning, have very different requirements than other business intelligence systems. In a traditional enterprise architecture, business events (e.g., a customer order, a supplier’s delivery delay) are communicated in a waterfall fashion from one business system to the next. AI agents, however, need to consume events immediately and independently of other systems. And this requires a different architectural approach to how events are shared and consumed, referred to as an event-driven architecture.

AI agents are often engaged in complex and closed-loop decisioning. They need to be notified of events in real time, cross-index other sources of business data for additional context, make decisions and initiate actions or other agent-enabled business processes. For example, a large customer order could be the precursor event for an inventory agent to check inventory levels and, if found to be low, re-order more inventory, possibly checking against demand forecast data and logistics data to reduce delivering time.

This kind of information consumption needs to happen in real time. At the same time, AI agents need to be selective in what they consume and have ready access to a wide range of resources. Event-driven architecture (EDA) platforms are designed to support these requirements for agentic AI systems.

In a traditional architecture, a workflow or business process may span multiple systems that are tightly coupled together to produce a specific outcome. In the case of a customer making an online purchase, for example:

Customer purchase (event) --> Order system --> Payment system --> Inventory system --> Delivering system --> Email system

Each step of the process requires the corresponding system or agent to be aware that something has changed and then take action: an invoice is created, the customer gets charged, the inventory system is updated, the delivery is scheduled and a confirmation email is generated.

In this arrangement, often called a “polling” architecture, AI agents have to poll each system to discover new events. An AI agent responsible for restocking might poll the inventory system every five minutes to see if inventory levels have fallen below a particular threshold. If inventory falls below that threshold one second after the last polling interval, the AI agent won’t know about it until the next poll. This limits agentic AI’s effectiveness by:

• Adding latency to decisioning.
• Using data that may be out of date.
• Consuming infrastructure resources even when nothing has changed.
• Creating more layers of complexity as AI agents scale.
• Making it a challenge to get a clear view of what’s happening in the business.

In an event-driven architecture (EDA), AI agents don’t poll systems for event information. Instead, events are published to the network and agents subscribe to these events, beginning a multi-stage process of subscribe, decide and act.

In the subscribe stage, the agent subscribes only to the types of events that interest it. When a relevant event is published, the agent is immediately notified. Next, the agent decides what to do about that event. Often, the agent may need additional context from other systems before making a decision. Finally, the agent acts. This action, in turn, may generate a new event that is acted upon by other event-driven AI agents.

These event-driven interactions require infrastructure capable of routing, filtering and distributing business events at enterprise scale. Solutions such as SAP Integration Suite, provide this event backbone, allowing agents, applications and business processes to interact through real-time events rather than point-to-point connections and repeated polling.

An event-driven architecture does more than help AI agents react faster. It enables autonomous agents to scale reliably, adapt to change and work together effectively.

Contextual awareness is important to agentic AI decisioning. Consider an AI agent tasked with completing customer orders. Upon receiving an event that a delivery will be delayed, the agent may require additional context—Is surplus inventory available? Are there alternate suppliers? Is this a priority customer?—to determine the best response. EDA provides AI agents with this real-time context by:

• Delivering fresh information, not stale data from the last poll.
• Adding relevant context at the right moment.
• Providing better situational awareness by combining multiple event feeds.
• Accelerating decisioning.

In a traditional, tightly coupled system, AI agents communicate with each other by API call or point-to-point integration. This can become increasingly complex to manage as more AI agents are added. With EDA’s loose coupling arrangement, instead of a web of dependencies, event-driven AI agents operate independently. This makes it easier to scale and manage multi-agent environments.

In a tightly coupled architecture, the network of interdependencies means that, if one system fails, it’s likely to impact other systems. Even after the failed system recovers, updating and synchronising the other dependant systems can be a monumental effort.

EDA’s loosely coupled approach avoids this domino effect. The recovery is far more graceful, because events are buffered, not lost and simply consumed by the system when it becomes available again.

Event brokers are designed to operate at scale and low latency. Buffering absorbs any spikes in business operations (e.g., Black Friday purchasing volumes), ensuring that systems and agents don’t get overwhelmed.

Adding new systems and agents into your enterprise landscape is also more flexible and safer with EDA. In the past, adding a new system or agent required direct connection to other systems. This meant additional updates or configuration, perhaps even downtime and always with some level of business risk. With EDA, the new system or agent operates in a more isolated fashion and simply subscribes to the events that are relevant.

An event-driven pattern offers a flexible and scalable way to express a workflow where events trigger, coordinate or control the work performed by AI agents. A generic pattern for agentic AI could be:

Event occurs --> Agent is triggered --> Agent reason --> Agent acts --> New event occurs

Applying that pattern to a very large customer order, the flow might look like this:

1. Event: The CRM system publishes a large order received event.
2. Trigger: An inventory agent that subscribes to that event is activated.
3. Reason: The agent checks available inventory, existing commitments and replenishment schedules.
4. Act: The agent determines that additional inventory is needed and initiates a replenishment order.
5. New event: An inventory replenishment event is published.

Publish/subscribe (pub/sub) is an event-driven pattern that notifies subscribing AI agents of a business event as soon as it happens . This model is ideal for agentic AI as it:

• Allows for real-time, asynchronous data delivery to agents.
• Eliminates repeated and wasteful polling.
• Avoids direct or point-to-point integration complexity.
• Prevents modification and change risk to the publishing system.
• Scales seamlessly as more agents are added.
• Enables multiple agents to respond to the same event simultaneously.

In the pub/sub model, a single event can fan-out to multiple AI agents at the same time, enabling those agents to process the event independently and in parallel. An example of a fan-out pattern might look like:

By processing events in parallel, EDA’s fan-out pattern reduces latency. This also allows enterprises to add new agents without disrupting existing AI agents or systems.

An event broker provides a connective layer between AI agents and the applications where events occur, such as ERP, CRM, supply chain, HR, finance and e-commerce applications. At a high level, the event broker ensures that AI agents can consume event information from these applications in a reliable and consistent manner. At a more granular level, an event broker supports several capabilities:

• Real-time, low-latency access to event information
• Delivery of the right events to the right agents
• Geo-distributed scaling through the creation of an event mesh of multiple brokers (also assisting data sovereignty compliance)
• Support for monitoring and tracing
• Activation of further agent processing via new events

By now, the advantage of implementing an event-driven architecture for AI agents should be clear. But to make the business case stronger, there are three critical questions you need to ask:

1. What are the token costs?

Enterprises looking to agentic AI need to consider the potentially high cost of LLM tokens as a result of agentic AI processing. In a traditional polling architecture, token costs quickly add up as AI agents are constantly polling applications to check for event updates, many of which may not even be relevant to the agent. An EDA platform, such as SAP Integration Suite, advanced event mesh, reduces this downstream processing in two important ways:

• It replaces constant polling with selective subscriptions, so agents only consume tokens when there’s something worth processing.
• It provides dynamic filtering that further refines token traffic by using event context to route only relevant events to agents.

2. Will adding new agents disrupt our existing systems?

As we’ve already noted, tightly coupled systems present inherent complexities when new agents or systems are added. A loosely coupled EDA system like SAP Integration Suite, advanced event mesh allows enterprises to easily add new agents with little or no additional integration. The result is lower integration costs, added flexibility and seamless scalability for AI agents.

3. Do they support dynamic routing/filtering and/or multi-cloud/hybrid deployments?

It’s not uncommon for an enterprise to have SAP applications in one cloud provider, AI agents in another, multivendor SaaS applications in a third location and additional applications on premises or at the network edge. A platform that can’t span those environments seamlessly will lead to yet more silos and additional integration complexity. SAP Integration Suite, advanced event mesh features distributed event brokers, dynamic routing and fine-grained filtering to support event-driven workflows across SAP and non-SAP applications in multi-cloud, on-prem and edge environments.

As enterprises adopt AI agents to automate increasingly complex decisions, the challenge becomes delivering the right business context at the right moment. AI agents are most effective when they can detect meaningful events, understand their implications and act in real time. Event-driven architecture provides the foundation for this model by replacing inefficient polling with timely and relevant event distribution, enabling agents to make faster, more informed and more coordinated decisions across the enterprise.

SAP Integration Suite, advanced event mesh helps organisations build that foundation by connecting applications, systems and AI agents through a scalable event backbone. With capabilities such as publish/subscribe messaging, dynamic routing and filtering, distributed event brokers and support for hybrid and multi-cloud environments, organisations can create event-driven automation that grows with their agentic AI initiatives.