Parallel Agents

This is a pattern for running up to 15 agents in parallel without destroying your context window. The key insight: don't use TaskOutput to check on agents because it dumps their entire output into context. Instead, have each agent append a completion line to a shared status file, then read that file to see who's done. Agents run with run_in_background: true and write "COMPLETE: task-name" when finished. For tasks that need more than just confirmation, each agent gets its own directory under .claude/cache/agents/. It's simple but solves a real problem if you're orchestrating batch work like backfilling data across providers or running parallel research tasks.