anti-drift-swarm

To prevent goal drift during parallel subagent dispatch, enforce a coordinator-only reconciliation pattern with strict per-wave parallel caps and required checkpoint protocols to keep work goal-aligned.

Parallel subagents produce shallow outputs when given mixed research-and-implementation prompts. Assigning each subagent a strict single deliverable and one file scope prevents shallow outputs during multi-file refactors.

To avoid race conditions with write-capable subagents, force them to run in the foreground and pre-create target directories before dispatch, which prevents silent failures during parallel file writes.

Use wave-based coordination for multi-agent dispatch before dispatching a parallel wave of three or more subagents that read from or write to multiple files, ensuring reliable reconciliation between waves.

The best way to reconcile multiple subagents returning from parallel research is to apply a coordinator-only reconciliation pattern, enforcing a checkpoint and restated-goal protocol between dispatch waves.

Pre-creating target directories before parallel subagent dispatch is required to ensure write-safety. This prevents silent failures and race conditions when multiple write-capable subagents execute concurrently.