What problem does it solve? Single-agent autonomous hill-climbing for LLM experiments gets stuck in local optima, repeats the same failure patterns, and leaves most of the search space unexplored. This Skill runs N agents in parallel, each with a different exploration strategy, so overnight research covers a wider search space and escapes local optima. ## Core Features & Use Cases - Multi-agent parallel exploration: Spawns N agents (default 3) in isolated workspaces under .research/<slug>/swarm/agent_<id>/tinker/, each running the autonomous tinker loop with its own strategy. - Five exploration strategies: depth-explore (model scale), lr-explore (optimization), arch-explore (structural ideas like RoPE/RMSNorm/SwiGLU), batch-explore (throughput vs gradient noise), and random-restart (uniform sampling to escape local optima). - File-based orchestration: swarm_orchestrate.sh periodically aggregates each agent's BEST.json into SHARED_BEST.json and SWARM_RESULTS.md using flock-based locking and atomic writes, with optional cross-pollination via a shared best_train.py snapshot. - Use Case: Launch a 3-agent overnight swarm on a language-model pretraining task where one agent tunes depth/width, another sweeps learning rates, and a third tries architectural variants, then read the aggregated SWARM_RESULTS.md to find the global best configuration. ## Quick Start Initialize a 3-agent research swarm for my project slug and start each agent's autonomous tinker loop with its assigned exploration strategy.
