Sub-Agents

Spawn parallel Claude Code processes, each with its own compute, to divide work across multiple GPUs. Each agent works autonomously and reports findings back to the parent via structured reports. This is a general orchestration pattern — use it for parallel debugging, parallel experiments, or any work that benefits from multiple agents.

IMPORTANT: Deploy the sandbox/experiment app ONCE, then spawn agents against it. Do NOT use modal run per agent — concurrent modal run processes conflict and kill each other.

Quick Reference

Deploy Once, Call Many

# Deploy ONCE (shared by all agents)
modal deploy modal-gpu-dev/tools/gpu_sandbox.py