Sub-Agent Delegation

Permissions

• NEVER spawn without explicit permission
• ASK first: "I've identified [TASK] for sub-agent delegation. Should I spawn one?"
• Explain WHY before requesting

When to Delegate

• GPU kernel optimization with iterative benchmarking
• Numerical correctness verification across test cases
• Performance profiling and analysis
• Parallel investigation of independent code paths
• Long-running validation suites

Patterns

• Parallel: Optimize independent kernels simultaneously (attention to A, MLP to B)
• Correctness First: Make tests pass before performance
• Incremental: Iterate until target speedup or report blockers

Kernel Optimization Template

Optimize [OPERATION] in [FILE].
Context: [current impl], [bottleneck source], [target HW: 3090/H100], [use case: train/inference]
Requirements:
1. Implement with Triton/CUDA
2. Verify: torch.allclose(atol=1e-5, rtol=1e-5), gradients match autograd
3. Benchmark: warmup=10, bench=100, report min/max/mean/std us
4. Scales: (1,128), (8,512), (32,2048)
Report: correctness status, perf table (scale, baseline_us, opt_us, speedup), memory

Workflow

Setup -> Develop -> Verify -> Benchmark -> Report

Requirements

• Report measured numbers, never estimates
• Include methodology (warmup, iterations, sync)
• Flag regressions immediately