{
  "task": "Implement TRM × Titans-MAG architecture",
  "requirements": {
    "model_type": "transformer",
    "parameters": "25M target",
    "features": ["sliding_window_attention", "long_term_memory", "ACT"],
    "framework": "pytorch",
    "constraints": {
      "vram": "6GB",
      "inference_speed": "real-time"
    }
  },
  "existing_code": {
    "files": ["model.py", "config.py"],
    "status": "partial_implementation"
  },
  "reference_materials": {
    "papers": ["TRM.pdf", "Titans-MAG.pdf"],
    "implementations": ["reference_model.py"]
  }
}

{
  "status": "implementation_complete",
  "deliverables": {
    "code_files": [
      {
        "path": "src/model/titans_mag.py",
        "description": "Titans-MAG backbone implementation",
        "loc": 350,
        "tested": true
      }
    ],
    "tests": [
      {
        "path": "tests/test_model.py",
        "coverage": 95,
        "all_passing": true
      }
    ],
    "documentation": {
      "architecture_diagram": "docs/architecture.md",
      "usage_examples": "examples/train.py"
    }
  },
  "performance_metrics": {
    "parameter_count": "25.6M",
    "inference_time": "45ms (GPU)",
    "vram_usage": "5.2GB"
  },
  "validation": {
    "unit_tests": "48/48 passing",
    "integration_tests": "12/12 passing",
    "manual_verification": "Forward/backward pass successful"
  }
}

from claude_agent_sdk import ClaudeSDKClient, ClaudeAgentOptions
import asyncio

async def execute_ml_expert(context: dict):
    """Spawn ML implementation specialist agent."""

    # Load specialist agent prompt
    with open('agents/ml-expert-specialist.prompt', 'r') as f:
        specialist_prompt = f.read()

    # Configure agent with write permissions (plan mode for safety)
    options = ClaudeAgentOptions(
        model='claude-sonnet-4-5',
        system_prompt=specialist_prompt,
        permission_mode='plan',  # Show intent before editing
        allowed_tools=['Read', 'Write', 'Edit', 'Bash', 'Grep'],
        setting_sources=['project']
    )

    client = ClaudeSDKClient(options)

    try:
        await client.connect()

        # Format task for agent
        task = f"""Implement ML solution:

Requirements: {context['requirements']}

Existing code: {context['existing_code']}

Reference materials: {context['reference_materials']}

Deliver production-quality implementation with tests and documentation."""

        await client.query(task)

        # Collect implementation results
        results = []
        async for message in client.receive_messages():
            if message.type == 'assistant':
                results.append(message.content)

        return parse_implementation(results)

    finally:
        await client.disconnect()

User: "Implement the TRM × Titans-MAG architecture from these papers with 25M parameters"

Skill gathers:
- Paper PDFs with architecture details
- Target parameter count constraint
- PyTorch as framework
- GPU memory constraint (6GB)

Agent implements:
- Titans-MAG backbone (sliding window attention, LMM, MAG gate)
- TRM wrapper (multi-pass reasoning)
- ACT head (adaptive computation)
- Full model integration
- Unit tests (95% coverage)
- Usage examples
- Architecture documentation

Deliverables:
- src/model/titans_mag.py (350 LOC)
- src/model/trm_wrapper.py (180 LOC)
- src/model/act_head.py (120 LOC)
- src/model/full_model.py (200 LOC)
- tests/test_*.py (48 tests, all passing)
- docs/architecture.md

User: "The ACT head has variance=0 issue. Add diversity regularization."

Skill gathers:
- Current ACT head implementation
- Diagnosis from ml-training-debugger skill
- Recommended fix (diversity loss)

Agent implements:
- Modify compute_act_loss() method
- Add diversity regularization term
- Update docstrings
- Add test for variance>0
- Verify training runs without warning

Deliverables:
- Modified src/model/act_head.py
- New test: tests/test_act_diversity.py
- Validation: Warning eliminated in training

User: "Model inference is too slow. Optimize for real-time performance."

Skill gathers:
- Current model code
- Profiling results
- Performance requirements (< 100ms)

Agent optimizes:
- Enable gradient checkpointing
- Fuse operations where possible
- Use torch.compile() for JIT optimization
- Optimize tensor operations
- Add caching for repeated computations

Deliverables:
- Optimized model code
- Performance benchmarks (45ms → 28ms)
- Memory usage reduced (6.2GB → 5.2GB)
- All tests still passing