Параллельный запуск агентов без переполнения контекста

orchestration

parallel-agents is a Claude Code skill that implements a pattern for running multiple agents in parallel without bloating the orchestrator's context. Instead of using TaskOutput — which returns full agent output and inflates context — agents append a single completion line to a shared status file using `>>`, allowing concurrent writes without conflicts. Up to 15 agents can run simultaneously per batch, each launched with `run_in_background: true` and a self-contained prompt that includes all necessary context, since agents do not share memory. Results are stored either as a single `.claude/cache/<batch>-status.txt` for simple confirmation or in per-agent directories with `output.md`, `artifacts/`, and `status.txt` for research-style tasks. Progress is checked via `cat`, `wc -l`, or `tail -f`, making this skill well-suited for bulk data processing, parallel exploration, and provider backfills.