We are fast approaching the point where distributed computing systems will be able to think on their feet, adapt to major changes in seconds, and silently solve problems without human intervention. This is the promise of agentic artificial intelligence (AI), a technological paradigm in which autonomous agents powered by large language models (LLMs) may soon organize and reorganize themselves to operate critical systems and applications entirely on their own.
Not just another technical upgrade, the leap to agentic AI promises to fundamentally redefine how computing systems are built and how software functions in the real world.
While the unusual complexity of agentic AI may prove difficult to manage, its tremendous potential to accelerate missions and disrupt industries underscores the need for organizations to build their awareness of AI that exercises agency. When AI can work by itself to implement directives and strategies without specific prompts and inputs, what new possibilities come into focus—and what are the risks and tradeoffs?
Distributed systems date to computing’s earliest days, when low-level mechanisms such as socket calls managed communication between a system’s different parts. In these early systems, each component had to know the exact memory identifiers, or “addresses,” of the other components it interacted with, so communication protocols were rigidly defined. This tight coupling made the systems brittle and difficult to modify or scale. As applications became more complex, any change to a single component potentially required redeployment of the entire system, leading to increased risk of bugs while extending development cycles.
In response to the limitations of older, monolithic ways of building applications, the microservices architecture emerged. This paradigm breaks applications down into smaller, more loosely coupled services that can be developed, deployed, and scaled separately. Each microservice typically encapsulates a specific business function and communicates with other services through well-defined application programming interfaces (APIs). This modular approach enhances scalability and agility to some extent, allowing teams to iterate on individual services without affecting the entire system. However, microservices architecture also entails significant drawbacks:
Fast-forward to the present day and the rise of generative AI (GenAI). While GenAI doesn’t yet have a human-level “brain,” it demonstrates capabilities that often serve as proxies for human reasoning and flexibility. Unlike tightly coupled systems that rely on predefined rules and centralized control, agentic AI systems use autonomous or semiautonomous generative agents capable of dynamic, real-time interactions. Now think of each microservice or application that needs to interact with another as an agent rather than merely a piece of software.
The emerging hypothesis—now becoming increasingly plausible— is that these agents will soon be able to communicate with each other dynamically and flexibly, without relying on fixed, predefined interactions. Historically, rigid system interactions made applications brittle, often breaking when evolving requirements or scaling added complexity. In contrast, agents leverage context, past experiences, and reasoning to respond effectively to unexpected or novel situations and tasks. This humanlike resilience enables agents to act quickly and precisely, even in uncertain or unpredictable environments, without requiring step-by-step instructions.
Even though a microservices framework with representational state transfer (REST) APIs introduces a degree of looseness by embracing the statelessness of Hypertext Transfer Protocol (HTTP), they do not make large software systems immune to becoming more tightly coupled over time. As large software systems mature and grow, they face requirements that naturally motivate tighter coupling, such as implicit dependencies; shared data models; cross-cutting features, such as logging and authentication; and backward compatibility to legacy code. With agents, the communication is entirely flexible, allowing them to interact and autonomously adapt, experiment, test, and optimize, working toward innovation without any manual intervention or reprogramming to implement new features or problem-solving strategies.
The ability of agentic AI to self-organize and adapt creates new use cases with relevance across industries. For example, GenAI has evolved from using single LLM applications to leveraging systems of specialized agents working in concert. Today, agents often have distinct roles, such as job planning and dispatching, information retrieval, content review, or output optimization. They typically operate within orchestrated frameworks to collaboratively solve complex, multistep problems.
Common design patterns for agentic AI systems include:
These agentic capabilities allow organizations to scale operations, improve efficiency, and tackle more sophisticated challenges by integrating multiple specialized agents or tools into a cohesive system. Here are two examples of how reference architectures can support different use cases.
Agentic AI can revolutionize operational planning by enabling more dynamic and responsive decision-making processes. For example, a multi-agent system for operations planning could involve a centralized commander agent that orchestrates and dispatches tasks to specialized worker agents. These worker agents have the autonomy to use various tools to retrieve required information and execute specific tasks. The system incorporates a feedback loop that allows for collaborative critique and compliance checks on drafted plans, ensuring that the final course of action is both effective and compliant with strategic objectives.
Figure 1: Reference architecture for operations planning system
In a document authoring use case, a multi-agent system involves a supervisor agent that classifies user requests into predefined intents and initiates stepwise workflows. These workflows incorporate the interaction and collaboration of different document writing agents, each responsible for specific aspects of creating documents. The system follows sequential logic flows that mimic human writing behaviors, ensuring that the final document is coherent and meets the specified requirements.
Figure 2: Reference architecture for document coauthoring system
Agentic systems offer federal agencies transformative potential by improving efficiency and addressing complex challenges. For instance, in disaster response and emergency management, a centralized commander agent could orchestrate worker agents to retrieve real-time weather data, review social media posts with pictures and videos for damage assessment, analyze geospatial information for evacuation routes, and coordinate emergency resource allocation. This streamlines decision making under tight deadlines, reduces the risk of human error, and enhances situational awareness—all of which are critical in high-pressure scenarios.
In compliance-heavy environments, agentic document coauthoring can accelerate policy drafting and ensure thoroughness. By drafting technical writing, reviewing for compliance, and contributing domain-specific insights, these systems reduce manual workload and enable faster turnaround times for critical documents. While risks such as reliance on incomplete data or unexpected agent behavior exist, users of these systems can mitigate these risks through oversight, thoughtful engagement, and review of outputs to ensure they are reliable and aligned with mission-critical objectives.
While agentic AI redefines distributed systems by enabling highly flexible, self-organizing architectures with vast potential, new challenges and tradeoffs also arise. The increased flexibility and autonomy agents offer will simultaneously create exponentially greater complexity. Designing, managing, and ensuring the reliability of these systems will mean harnessing sophisticated approaches and frameworks to tame chaos and address multiple areas of risk, described in the following sections.
Agentic AI holds significant potential, yet its widespread adoption may be hindered by high costs. Here’s why: Each interaction between agents incurs a cost associated with running inference. As the number of agents increases and their autonomy grows, computational expenses proportionally rise—and today we don’t have certainty, or even reliable estimates, of the cost and extent of this dynamic interaction. Researchers at Princeton University found that current research into AI agents often focuses narrowly on accuracy, while neglecting analyses of cost control measures. However, the cost of hardware and computation tends to decline over time. While it may be very expensive today, costs could become reasonable for most enterprises within the next two years.
The resource-intensive behavior of agents constantly “thinking” or negotiating tasks creates significant operational challenges. Unlike traditional microservices that execute predefined tasks efficiently and predictably, autonomous agents may continuously process real-time data, make decisions, and adapt to changing conditions. As the number of agents in a system increases, the complexity of their interactions grows. Scalable solutions must efficiently handle the increased load without compromising performance.
Training agents with relevant data presents another challenge. Autonomous agents are most effective when trained on certain tasks, actions, or reasoning. To train these agents on a specialized domain or new task often requires huge amounts of high-quality data. For example, for tasks like transcription, summarization, and sentiment analysis, extensive datasets must represent real-world diversity and nuances. These tasks require an understanding of context, tone, and language subtleties, which vary across domains. Exhaustive training to handle these variations underlies successful deployments.
The loose coupling in agentic AI systems introduces significant security and stability risks. Autonomous agents evolve and adapt their behavior over time. This flexibility, while offering obvious advantages, means agents may develop in unexpected ways, exposing vulnerabilities or behaving unpredictably. Concerns also arise from the potential of agents communicating in formats humans can’t read and understand. As agents optimize interactions, they may develop truly opaque communication protocols beyond natural language, markedly complicating the process of auditing and understanding their decisions and actions. If agents decide that communicating in binary is more efficient, how will humans ever find their way back into the conversation? To ensure trust and safety in these systems, new tools to log, monitor, and audit agent interactions will likely emerge.
As autonomous agents take on more responsibilities, determining who or what is accountable for their actions becomes a complex issue. Traditional accountability frameworks, which rely on human oversight, may not be sufficient. If organizations don’t handle data responsibly and in compliance with privacy and security regulations, stakeholder trust may be lost. Are the decision-making processes of agents free of errors and bias? It may be difficult or impossible to know this in some cases.
The rapid evolution of agent-based systems will likely disrupt traditional software architectures as AI agents become increasingly capable of performing a broad spectrum of tasks. With applications beginning to incorporate more agents, common multi-agent architecture patterns will emerge, crystallizing a need to:
The adoption of agent-oriented architectures may signify a shift away from service-oriented and API-based architectures. While API integrations will remain necessary for certain types of transactions and functions (e.g., deterministic and recurring), agent-oriented architectures could lend themselves well to performing nondeterministic functions and tasks, especially for tasks that haven’t been tackled before. One day a team of agents may even collectively decide which GenAI algorithm to use for their next project, almost like picking the best “brain” for the task.
Given the utility and impact of agent-oriented systems, agentic components will become pervasive in all applications and software systems at some point in the foreseeable future. It’s likely that commercial software products will adopt agent orientation and provision access to not just APIs but to agents for complex interactions and functions. Leveraging these new capabilities to solve customer problems and successfully achieving product market fit will be critical to technology businesses.
Numerous challenges remain unaddressed regarding agent-to-agent communication protocols, mechanisms for agent discovery and registration, skill refinement based on environment feedback, and more. For example, current agent solutions are predominantly designed for human interactions, such as conversational AI in natural languages, rather than the machine-oriented communications typified by web API calls.
To safely and productively operationalize agentic AI systems, organizations will likely focus on basic risk mitigation strategies, with thorough testing and validation of agentic AI systems before deployment and the creation of protocols for human oversight and intervention when necessary. New research into data governance and ethical standards related to autonomy will help prevent misuse of agents or ethical breaches that create financial and reputational damage. Effective control systems will be urgently needed to monitor and regulate agents’ behavior, preventing them from deviating from their intended functions or engaging in harmful activities. Research on auditing platforms, traceability, and techniques will be required to log every interaction agents have with each other.
It’s also important to consider that the decentralized nature of agentic AI systems may be especially difficult to reconcile, at least initially, with the need for strict accountability within the federal sector. Ensuring that autonomous agents operate inside defined ethical and legal boundaries means building robust governance frameworks to begin with. The scalability of such systems must also be carefully managed to handle the vast amounts of data and complex interactions characteristic of federal operations.
There aren’t yet production systems with hundreds of agents working together at once, but agentic AI is a revolutionary leap in innovation, and it is moving so fast that organizations should start experimenting with these tools now, in a sandbox environment, to test their capacity to understand and govern agents effectively. By implementing the right prototyping, testing, and risk mitigation strategies, organizations will position themselves to harness the expansive technical benefits of agentic AI while ensuring that these systems operate in ways that are ethical and innovative, transformational, and safe.
leads Booz Allen's AI practice with a focus on ensuring leaders across federal missions achieve AI understanding, purpose-built solutions, and accelerated adoption.
Booz Allen’s director of generative AI, leads GenAI solutioning across the firm and helps teams create best practices for AI development and use.
Booz Allen's annual publication dissecting issues at the center of mission and innovation.
Want more insights from Velocity? Sign up to receive more stories about emerging technologies and the impacts they’re making on missions of national importance.
