efficient-multi-agent-research

Partitioning large codebase research across parallel sub-agents prevents context pollution by dispatching background investigators that write findings to disk. The main agent stays clean while a coordinator consolidates interim results into a single report.

The best way to audit function call sites is partitioning the review across multiple parallel sub-agents. A coordinator dispatches background investigators to examine different codebase sections and consolidates their written findings into one unified report.

Multi-agent research consolidation works by having a coordinator partition file reviews into tasks for background investigators. Each investigator writes findings to disk, and the coordinator consolidates these outputs into a single report, keeping the main context clean throughout the process.

You can use 4 to 5 parallel agents with run_in_background for codebase research before consolidation. This setup requires a central coordinator to partition the research workload, dispatch the agents, and consolidate their disk-written findings into a single report.

You should not use parallel sub-agents for codebase research when investigating fewer than six items. This approach requires a coordinator to partition work and dispatch background agents, making it excessive for small research workloads that do not risk polluting the main agent's context.