From AI models to agents: What makes a system “agentic”?
AI models and LLMs have made incredible progress in recent years, with uncanny gains in natural language processing to create and understand language. But the real shift happening right now is the development of modern AI agents to plan and execute end-to-end tasks. Consider the difference between earlier special-purpose AI systems, such as IBM’s Deep Blue chess engine in 1997, which could beat human champions handily and was focused on one specific task. Contrast today’s stochastic, tool-using AI agents that can plan and execute ambiguous end-to-end tasks like “book a trip to San Francisco next week under $1,000” that require many steps and autonomous decision-making during the process.
I recently became a Claude Certified Architect, Foundations, and that process was invaluable in helping me understand the structure, complexities, and opportunities of agentic AI. I’m eager to share some of those core principles.
Let’s start by defining how agentic AI architecture compares to traditional AI. A key feature of agents is their ability to use models, external tools, and memory to extend the basic capabilities of the large language models they’re built on. Agents work by planning goals autonomously, emphasizing the decomposition of complex workflows into multi-step tasks and, in many cases, enabling parallel execution. As we will discuss, this last aspect is one of the most compelling aspects of an agentic architecture: by using different agents with specialized skills and goals, the orchestrator can delegate tasks effectively and then combine the results. This is far more powerful than attempting to solve complex problems with a single user request.
Another way to think about agents is through the kinds of queries they can handle. An LLM can receive a single query and often produce an impressive and apparently thoughtful result. An agentic system, on the other hand, can receive an assignment to solve complex tasks, which require an iterative loop from perceive → reason → act → observe → update memory → repeat until the goal is met. Moreover, the agent is not confined in the LLM but connects to external systems to augment the agent’s knowledge and take actions. Depending on the user’s preference, many agents can operate with minimal human input.
From a business perspective, agentic AI systems became mainstream in enterprise discussions in recent years as organizations tried to move from chat-style copilots to autonomous workflows. As agent skills are exposed to company data at enterprise scale, their ability for task automation and handling of increased task complexity can significantly broaden their usage across many aspects of an enterprise.
Core agentic AI architecture components
Most agentic AI architectures can be decomposed into four core capabilities, though each type may be structured into one agent or multiple agents, depending on the specific nature of the problem being solved.
- An orchestration layer is responsible for managing the overall task and delegating components to sub-agents
- Perception tasks are oriented around understanding the assignment, intent, and processing what are often significant data sources to find relevant information
- Reasoning tasks consume the understanding of the task and the raw materials to reason and determine ideal courses of action
- Action-oriented tasks connect to outside services that publish their available capabilities, often through MCP (Model Context Protocol) servers, to enable outside actions.
In the coming sections, we will explore each of these components to understand how agentic AI systems are constructed and the many unique design aspects needed for successful and efficient agentic systems. This blog does not attempt to cover the myriad technical standards and interfaces being used to connect agentic systems or the complex environments of many distributed systems, but rather it provides an overall introduction to the critical tradeoffs and typical design approaches currently being employed.
Perception: How agentic systems ingest and understand the world
Perception is the “sensory” layer by which the agentic AI understands the world. To understand the context of the task at hand, agents need to read and understand data of a range of different broad types:
- Natural language inputs like chat inputs, email, and service tickets
- Non-structured document inputs such as specifications, web pages, PDFs, and more. Taking this type of high-density, unstructured content and turning it into understandable inputs to an agentic architecture is especially important as it enables AI agents to have the type of context humans would normally perceive.
- Structured data input such as SQL databases, sensor data, audio, video, or other object storage modalities.
Across all data types, the underlying challenge is converting them into a structured representation the reasoning engine can work with efficiently.
Orchestration and reasoning: The cognitive engine of AI agents
At the center of the agentic architecture is the orchestration and reasoning layer. The reasoning layer is the “brain” of an AI agent that chooses goals, decomposes them into plans, selects tools, and decides when a task is “done.” Today, LLMs (GPT‑4.1, Gemini Pro, Claude, LLaMA variants, etc.) are typical reasoning engines as of now, and form the core.
The difference between single LLM calls and an agentic approach is that the reasoning layer, including the overall orchestrator, breaks down the task into sub-goals, which decide when to delegate to other agents, when to ask humans, when to use tools, and when to terminate because the goal is not being met. In an agentic system, each sub-agent in a multi-agent setup is given a specific mission that is more focused than the overall mission. That specific mission comes with its own customized role, system prompt, and skills to help the agent to focus and achieve its specialized task more effectively.
Depending on the task, reasoning agents are carefully designed to ensure the scope of their work is achievable within the constraints of the LLM. Hallucinations and “forgetting” key facts can occur when maximum prompt lengths are reached, data is provided in an unstructured format, or the task is simply too large. In those cases, a multi-agent workflow may be needed where larger tasks are restructured into serial or parallel agents that work on a smaller subset. The complexity of multi-agent collaboration lies in managing the sub-tasks, which increases the coordination complexity of the orchestrator.
Memory: Short-term and long-term context for agentic AI systems
Agentic AI also necessarily relies on persistent memory to ensure that the agents can track their task progress, especially as agents operate over time, across multiple steps, sessions, or successive tasks. The simplest of these is the conversation history, which the agent framework must provide back to the reasoning layer since LLMs are fundamentally stateless. Structured context management is needed to ensure the conversation does not overflow as the task proceeds.
Separate from the short-term memory of the conversation history itself is long-term memory outside persistent storage, such as user profiles, past incidents, project histories, or domain-specific documents implemented via vector databases, knowledge graphs, or relational/SQL stores. Using long-term memory enables AI agents to create RAG (retrieval-augmented generation) pipelines that allow, for example, a customer support agent to retrieve past tickets and purchase history to personalize responses.
Action: Tool use, execution, and real-world effects
Fundamental to agentic AI is the ability to connect AI to decisions into real-world changes by calling tools, APIs, databases, or other agents, and by generating code, tickets, or configuration changes. As of this writing, tools are often exposed by MCP that publishes tools available for AI to use and provides a structured schema for input and output.
Tool use comes with important implications. As of today, overloading too many tools or providing overly complex schemas risks confusion, causing the reasoning engine to choose the wrong tools or overload context windows. Subdividing agents into specialized tasks and giving each the relevant tools for its sub-task is one strategy to ensure better success.
Equally critical is ensuring sufficient security so agents can only access approved tools, and that human oversight is used for areas of uncertainty, high risk, high financial implications, and so on. Error handling is important, both to make sure tools complete successfully and the agent gracefully manages the situation when it does not, such as managed retries, as well as providing guardrails to ensure tool calls are compliant with business logic or security policies.
Architectural types: Single-agent, multi-agent, and hybrid systems
Agentic systems can range in complexity. Today, single-agent systems can handle focused workflows, multi-agent systems are used for complex or large-scale problems, and hybrid setups blend human and AI agents across layers.
- Single-agent architectures: one AI agent with its own reasoning, memory, and tools running in a loop. A typical example is a customer support agent, internal assistant for documentation, and report-generation agents.
- Multi-agent architectures: These comprise multiple specialized agents (e.g., “Researcher,” “Planner,” “Executor,” “Reviewer”) coordinating via messages. Multi-agent ai design patterns range from hierarchical (coordinator + workers), decentralized peer-to-peer networks, and hybrids where leadership can shift based on task context. Each brings its own benefits and complexities.
Best practice is to start with single-agent systems and move to multi-agent only to address specific additional needs (e.g., the need for specialization, parallelization, verification), to avoid unnecessary complexity.
Outlook
Agentic AI can accelerate development of new capabilities and is being deployed widely across many industries and usage types. The science and capabilities of agents are evolving rapidly as more and more services publish MCP interfaces that allow agents to access them, whether for input or output actions. Sonatus is investing heavily in this area and delivering complex multi-agent systems to support use cases from pre-production test and validation through after-sales service. We are excited about the potential of this new type of compute and the potential it can bring to our customers.
