When Structure Pays

An agent system can spend extra computation in two ways: generate several candidates and choose among them, or add a parent process that delegates to workers. These articles compare both choices against a simpler one-agent baseline.

For teams choosing an agent design, the series shows when a shared checker makes extra samples useful, when coordination adds cost without a measured gain, and what another experiment must record before the result can generalize.

Start here: When Does Best-of-Five Sampling Pay for AI Code Generation?. Start with best-of-five sampling, where every candidate faces the same deterministic check. Then examine a paired supervisor-versus-solo replay that records quality, tokens, elapsed time, missing data, and uncertainty.