Basic Multi-LLM Workflows

Why it matters

Section 1.6 asks when the steps are known in advance. These three workflows are the fixed end of that choice, so you need to know what each one buys before you choose adaptive decomposition.

What to practise

Build chain, which feeds each prompt’s output into the next; parallel, which runs one prompt over several inputs in a thread pool; and route, which has the model explain its reasoning, pick a route in XML tags, and then sends the input to that route’s prompt. Run them on data extraction, stakeholder impact analysis and support-ticket routing. Watch what each pattern adds in calls and latency, and where a chain step’s error flows into every later step.

How it fits

In all three, the code fixes the possible steps in advance; even route only picks among prompts written beforehand. Compare them with 1.6.1, which chooses a fixed chain or adaptive decomposition depending on whether findings change the work, and with the file-then-integration chain in 1.6.2.