Sub-Agents (Parallel Agent Orchestration)

What it does

This skill provides an orchestration pattern to spawn and manage multiple autonomous Claude Code agents, each running in its own GPU-backed sandbox or as a lightweight experiment caller. It describes deployment best-practices (deploy once, call many), interactive SSH-driven sandboxes for debugging, and non-interactive batch experiment patterns. The skill includes tools for agents to report structured findings and for the parent process to monitor progress and read trajectories or summaries.

When to use it

Use this skill when you need to run parallel experiments (hyperparameter sweeps, model comparisons), concurrent debugging sessions across GPUs, or distributed data-processing agents that must report structured results back to a coordinator. It is ideal for workloads where isolation per-agent (separate sandbox/VM/GPU) improves reliability or throughput.

What's included

• Scripts: orchestration examples and CLI snippets in the skill body (has_references=true indicates companion docs).
• References: references/sub-agents.md with full orchestration guidance.
• Instructions: step-by-step deploy-once instructions, SSH sandbox flow, batch caller pattern, reporting and monitoring commands (agent_summarize.py, agent_report.py).

Compatible agents

Primarily designed for Claude Code / Claude-based subprocess orchestration; patterns apply to other LLM agent runtimes that support external process invocation and remote sandboxes (e.g., Sonnet, custom Claude CLI integrations).