AI Agentic Architectures: The Return of Divide and Conquer

Agent decomposition flow showing complex problems broken into prompts, solved by LLMs, executed by modular agents, and combined into a solution Divide and conquer never went away - agents just replaced the subroutines with reasoning engines.

At their core, AI agentic architectures are a modern manifestation of one of the oldest problem-solving paradigms in computer science: divide and conquer.

In traditional algorithms, the divide-and-conquer approach involves breaking a complex problem into smaller, more manageable sub-problems. These sub-problems are solved independently, and their solutions are then combined to address the larger issue.

AI agentic architectures adopt this same principle - just with a new toolkit.

Decomposition into Prompts: Instead of hardcoded logic, problems are reframed into natural language prompts. These prompts guide the reasoning process, splitting complex workflows into bite-sized tasks.

LLMs as Problem Solvers: Large Language Models act as the general-purpose solvers. They interpret the prompts, reason through ambiguous scenarios, and make decisions for the next steps.

Data and Action Spaces: Each “agent” operates within a defined data and action space. The data provides context, while the action space encapsulates what the agent can do - query APIs, trigger workflows, or fetch more context.

Agents as Modular Executors: Agents package these decomposed sub-problems into modular, reusable entities, each responsible for a specific facet of the larger workflow.

By stitching together agents - each solving a part of the puzzle - AI systems mimic the elegance of divide and conquer. But unlike traditional algorithms, these systems can reason, adapt, and evolve based on feedback.