Agentic AI
AI agents are autonomous entities that are capable of perceiving their environment, reasoning, and taking actions to achieve defined goals. Agents can interact with users, services, and other agents in complex, real-time environments. In generative AI, the core of an agentic system is usually a large language model (LLM), augmented with capabilities like retrieval (access to information), tools (interacting with its environment), and memory. The term agentic derives from agency (the ability to act independently, make decisions, and take actions within its environment). Agentic systems can have varying degrees of agency, ranging from making narrowly scoped actions to autonomously orchestrating themselves. When designing an agentic system, it's important to only increase the agency of the system when the task complexity requires it.
In this section of the Generative AI Lens, we'll discuss a few of the common agentic patterns we see in practice.
LLM-augmented workflows
LLM-augmented workflows are systems where code paths are largely deterministic, with certain steps augmented with LLMs to make decisions. An example is simplistic document processing system. It might classify an incoming document as simple or complex and use tool calls to route the document to the correct path. It has a low degree of agency, but is still an agentic system as a whole because it's making decisions through tool invocations.
Autonomous agents
The architecture of an autonomous agent is simple. These systems are typically built around an LLM that has been augmented with retrieval, tools, and memory orchestrated in a loop. This is commonly referred to as a ReACT (reason and act) loop. An agent is given a list of tools with instructions on how to use them and is orchestrated in a loop until the LLM is given a stop reason. These tools can be internal APIs, connections to data sources, or simple functions. Tools are often connected to the agent using the MCP protocol, and the LLM provides the agent with the reasoning capabilities needed to select the appropriate tool. The guidelines informing this reasoning capability are traditionally described in the system prompt, which should outline the goal, role, and other details the LLM needs to properly reason.
Hybrids
In practice, many agentic systems use both workflows and autonomous agents with varied levels of agency. An example is the plan and solve loop which has a high degree of agency.
First, we use an LLM to create a plan and a task breakdown. The plan and task list are informed by the agent's system prompt, which broadly describes how the LLM should reason on the agent's behalf.
That task list gets fed into a queue and we use a series of ReACT agents to resolve those tasks.
When the task list is complete, we use another call to an LLM to decide if we have enough to return to the user or need to add more tasks to complete the users request.
This pattern could be scaled out using queues, auto-scaling, and asynchronous executions.
Conclusion
As we navigate the evolving landscape of generative AI, the agentic paradigm represents a fundamental shift in how we architect intelligent systems. From simple LLM-augmented workflows to fully autonomous agents, these patterns enable us to build solutions that can perceive, reason, and act within their environments with varying degrees of agency. The key insight for practitioners is that agency exists on a spectrum. Our architectural choices should align the level of autonomy with the complexity of the problem at hand.
As agentic AI continues to mature, we anticipate these patterns will evolve and new architectures will emerge. The principles of well-architected design (reliability, security, performance efficiency, cost optimization, sustainability, and operational excellence) remain paramount as we build systems that act on our behalf. The future of generative AI is increasingly agentic, and mastering these architectural patterns today prepares us for the autonomous systems of tomorrow.
