The five workflow patterns

On top of the four agentic patterns, the survey describes five workflow patterns: structured ways to organize how LLM calls and agents are wired together. These come from the broader agent-building literature (the survey cites Anthropic's "Building Effective Agents" and LangGraph tutorials). Different patterns suit different task complexities.

Prompt chaining

A complex task is decomposed into a sequence of steps, where each step builds on the output of the previous one. Breaking the problem into simpler subtasks improves accuracy. The tradeoff is added latency, because the steps run in sequence rather than at once.

Routing

An incoming query is classified and directed to a specialized process or tool best suited to handle it. Rather than treating every query the same, a routing step sends a database question to a database tool and a web question to a web search. This improves both efficiency and relevance. The single-agent "router" architecture in Chapter 3 is built entirely around this pattern.

Parallelization

A task is split into independent parts that run concurrently, and the results are combined. This is the key to speeding up read-style work, because independent subtasks do not have to wait for each other. It only helps when the parts are genuinely independent, echoing the read-versus-write lesson from the multi-agent topic.

Orchestrator-workers

A central orchestrator agent dynamically breaks a task into subtasks, delegates them to worker agents, and synthesizes their results. Unlike simple parallelization, the orchestrator decides the decomposition at runtime based on the query. This is exactly the pattern behind Anthropic's research system from the multi-agent topic.

Evaluator-optimizer

One agent generates a result while another evaluates it against criteria, and the feedback drives iterative refinement until the output is good enough. This is reflection turned into a two-role loop, and it is the engine behind the "corrective" RAG architecture in Chapter 3.