Agentic Reasoning: LLMs that Plan, Act, and Learn

WHY IT EXISTS Large Language Models (LLMs) excel at reasoning in closed-world, static scenarios but struggle with open-ended, dynamic problems. They lack the ability to interact with an environment, learn from feedback, and adapt their strategy. Agentic reasoning was created to bridge this gap, enabling LLMs to move from passive text generation to active problem-solving.

THE MENTAL MODEL Think of an agentic LLM not as a chatbot, but as a digital worker operating in a cycle. You give it a high-level goal, and it creates a plan, executes steps using tools (like a web browser or code interpreter), observes the results, and adjusts its plan based on that feedback. It's a continuous loop of thought, action, and observation until the goal is achieved.

HOW IT WORKS An agent operates in a reasoning cycle. First, it PLANS by breaking a complex goal into a sequence of smaller, manageable sub-tasks. Second, it ACTS on a sub-task, often by using an external tool like calling an API or running a search query. Third, it OBSERVES the outcome of its action. This new information is used to LEARN and refine the next iteration of the plan. This process can be orchestrated at runtime with structured prompts (in-context reasoning) or baked into the model through fine-tuning on successful action sequences (post-training reasoning).

WHEN TO USE IT Use agentic frameworks for complex, multi-step problems requiring interaction with external systems. This is ideal for tasks like autonomous research, where an agent must browse the web and synthesize findings, or for automated software engineering, where it might write, test, and debug code. It excels in dynamic environments where the path to a solution is not known in advance.

WHEN NOT TO USE IT Avoid this complexity for simple, single-shot tasks where a direct prompt-and-response is sufficient, like summarizing a fixed piece of text. The overhead of planning, tool use, and state management is unnecessary for problems that don't require interaction or adaptation. It is also risky in high-stakes domains without robust human oversight and governance.

ONE CANONICAL EXAMPLE A research agent is tasked to “summarize recent breakthroughs in solid-state batteries.” It first plans to search academic databases. It uses a search tool to find papers, observes the results, and identifies key themes. It then refines its plan to read the abstracts of the most relevant papers, uses a tool to access them, synthesizes the information, and finally generates a report. If a search fails, it adapts its plan by trying new keywords.