Coordination patterns

Five foundational patterns, simple to complex, form the common vocabulary for composing LLM calls into systems: prompt chaining, routing, parallelization, orchestrator-worker and evaluator-optimizer. The wrong pattern fails regardless of model quality; the right one works with even modest models.

Same five, at working resolution: each pattern is a claim about the task. Predefined path? Chain. Distinct input categories? Route. Independent subtasks or value in multiple perspectives? Parallelize. Subtasks unknowable upfront? Orchestrate. Clear criteria plus measurable gains from refinement? Evaluate and loop.

Why coordination decides outcomes

It is well understood and documented by now that most production failures stem from mismatches between task requirements, coordination patterns and information flow. Not model quality. Nor does performance increase monotonically with more agents. Instead, communication overhead grows, context windows overflow and systems begin self-interfering without clear structure. As such, architecting agentic AI is more about which coordination pattern matches the task at hand.

The five patterns

Anthropic's research identified five foundational coordination patterns that have become the common vocabulary for agent architecture, explored here from simple to complex.

Prompt Chaining refers to a pattern where each LLM call processes the previous output in sequential steps with programmatic gates verifying progress between steps. The path is predefined. It is a pattern similar to Plan-and-Execute with the distinction that in this approach every step is a reasoning one and thus an LLM call.

Routing describes an approach during which the agent classifies the input and then directs it to a specialized handler. Different categories get different prompts, tools and even different models. Easier tasks go to cost-efficient models while harder ones are tackled by frontier ones. This approach is used when there are distinct input categories that benefit from separate handling.

Parallelization is expressed in two variants. Sectioning, within which the task is broken into independent subtasks that run simultaneously, and voting, in which the same task runs multiple times for diverse outputs and then results are aggregated. The pattern is used when speed matters or when multiple perspectives increase confidence.

Orchestrator-Worker describes the approach where a central LLM dynamically breaks down tasks, delegates to workers and synthesizes results. The key difference from parallelization is that subtasks aren't predefined. They emerge based on the specific input. This pattern is used for complex tasks where you cannot predict the subtasks in advance.

Evaluator-optimizer is the pattern within which one LLM generates and another evaluates and provides feedback, looping until quality meets a threshold. When there are clear evaluation criteria and iterative refinement measurably improves results, this approach fits.