coordinate-subagents

Coordinate subagents by decomposing tasks into independent units, running them in parallel, and applying token budgets to each. This prevents sequential bottlenecks and reduces token consumption compared to running subagents sequentially or with redundant context passing.

Use voting when subagents must reach consensus on critical or irreversible actions. Voting ensures reliability by aggregating independent judgments, reducing the risk of unreliable outputs from a single subagent.

Implement strict format constraints in subagent coordination by specifying required output schemas upfront and validating results against them. This prevents malformed outputs and ensures downstream processing reliability.

Yes. Right-sized models means assigning appropriate model capability to each subagent task—simpler tasks use smaller models, complex reasoning uses larger ones. Combined with parallel calls and minimal context, this improves token efficiency and latency.

Extreme decomposition breaks complex tasks into minimal, independent units so subagents can run in parallel without redundant context or blocking dependencies. This eliminates sequential bottlenecks and reduces overall token consumption.