
from deepswarm112
Orchestrate large-scale parallel AI workers for batch and multi-turn tasks with auto-calculated worker counts, stagger, and tiered model delegation for cost-eff
DeepSwarm runs and manages N parallel worker processes to execute large batch or multi-turn API tasks. It auto-calibrates worker count, stagger delays, and batch sizing to maximize throughput while avoiding rate limits. Supports tiered delegation where an orchestrator (frontier model) plans and cheaper workers execute at scale.
Use for long-running generation, translation, summarization, or classification pipelines that benefit from parallelization (call durations >10s), or when you need crash-resilient checkpointing and high throughput across many seeds. Not intended for short synchronous calls or tightly coordinated inter-worker tasks.
Useful for agents that can schedule and monitor long-running background jobs, pipelines, or developer-run CLI workflows (Hermes-style orchestrators, tmux-agent orchestrators, and batch-processing tools).
DeepSwarm is a parallel worker orchestration skill for batch API tasks with auto-calculated worker counts and stagger delays. Only seed.py ran successfully out of the box; the other three scripts require runtime arguments and a task.yaml config file. Both swarm.py and worker.py hardcode reading API keys from ~/.hermes/.env, which is a security concern. The SKILL.md is well-structured with clear instructions and useful pitfalls section, but the scripts lack type hints and have minor inconsistencies.
pyyamlrequestsHardcoded credential file path (~/.hermes/.env) in two scripts is the main security concern. Not malicious but bad practice — should use env vars or a configurable path. The skill is genuinely useful for batch API orchestration, and the auto-optimization logic is well thought out.