Doing things at once
By the end: Use sectioning and voting appropriately, and recognise when parallelism is false economy.
Parallelism helps in two distinct shapes, and confusing them wastes money.
Sectioning splits work with genuinely independent parts and runs them together. Ten documents to summarise, thirty files to review, five products to describe. Latency collapses to roughly the slowest one, and each call gets a clean, small context rather than one call trying to hold everything.
Voting runs the same task several times and compares. It buys reliability where being wrong is costly and judgement varies: does this code contain a vulnerability, does this content break the policy. Three runs agreeing is meaningfully stronger evidence than one, and you can tune how many must agree to how much you fear a false negative.
Within a single agent turn, a model can request several independent tool calls at once. Running those concurrently instead of in series is free latency, and it is a common oversight in first implementations.