README.md

Architect

The optimization framework that powers Engram and all baseline methods. It provides the CLI, task loading, LLM interface, and base classes for optimization methods.

See the examples README for how to run each method.

---

Directory Structure

Architect/
├── main.py          # CLI entry point
├── task.py          # Task class: wraps a task prompt + evaluator
├── types.py         # Shared type definitions
├── utils.py         # Logging and directory utilities
├── methods/         # All optimization method implementations
│   ├── common.py    # OptimizationMethod base class
│   ├── agentic_handoff.py   # Engram
│   ├── evolution.py         # Evolution of Heuristics
│   ├── single_agent.py      # Single-agent baseline (Glia)
│   ├── ...                  # Other methods
│   ├── deepagents_utils/    # System prompts and agent configs
│   └── handoff_utils/       # Archive and Research Digest logic
├── llm/             # LLM interface (OpenAI API) and prompt compilation
└── openevolve/      # Adopted from github.com/algorithmicsuperintelligence/openevolve

---

Adding a New Method

1. Create a new file in methods/ (e.g., methods/my_method.py).

2. Subclass OptimizationMethod from methods/common.py and implement the optimize() method:

from Architect.methods.common import OptimizationMethod

class MyMethod(OptimizationMethod):
    def __init__(self, task, model, results_dir, debug=True, **kwargs):
        super().__init__(task, model, results_dir, debug)
        # your init here

    def optimize(self):
        # Use self.task, self.architect, self.log_dir, etc.
        # Return a results dict with at minimum {"best_solution": {"score": ..., "code": ...}}
        pass

3. Register the method in main.py by adding it to the method dispatch logic.

4. Create an example script in examples/ following the pattern of existing scripts (see examples/handoff_example_usage.py).