Basic Multi-LLM Workflows

Why it matters

Before granting an agent autonomy, see how far a fixed workflow gets you: in these three, the code keeps control of the path.

What to practise

Write a chain function that feeds each step’s output into the next prompt, and use it to turn raw text into a formatted table. Write a parallel function that runs one prompt over several stakeholder groups at once with a thread pool. Write a route function that asks Claude for its reasoning and a <selection> in XML, then sends the support ticket to that team’s specialized prompt. Watch what each trades in cost or latency, and what route does when the selection matches no route key.

How it fits

All three are workflows in 1.3.1’s sense: the host owns the execution path. Match each to the task dependencies it suits, as 1.3.2 does; 1.3.2 also gives a router an unresolved-input path, which this one lacks. 1.3.4 is the record that defends picking the simplest of these over an agent.