Choosing Between LangGraph, OpenAI Agents SDK, Google ADK, and Microsoft Agent Framework

This repository provides an offline deterministic decision matrix for evaluating agent framework trade-offs under explicit production constraints.

It is an offline, deterministic decision matrix. It scores framework records against explicit project constraints. It is not a benchmark, does not call external services, and does not claim that one framework is right for every context.

What It Does

• Loads framework capability records from data/frameworks.yml.
• Loads scenario constraints from data/constraints.yml.
• Validates the schema before scoring.
• Applies weighted scoring and minimum capability gates.
• Prints a ranked result for the selected scenario.

The .yml files intentionally use JSON-compatible YAML so they can be parsed with Python's standard json module. This keeps the project dependency-free and deterministic.

Source Boundaries

The framework notes are conservative summaries based on official documentation verified on 2026-06-03:

• LangGraph reference: https://reference.langchain.com/python/langgraph/overview
• OpenAI Agents SDK guide: https://platform.openai.com/docs/guides/agents-sdk/
• OpenAI Agents JS guide: https://openai.github.io/openai-agents-js/guides/agents/
• Google ADK: https://adk.dev/
• Google ADK docs redirect: https://google.github.io/adk-docs/
• Microsoft Agent Framework: https://learn.microsoft.com/en-us/agent-framework/

The scoring values are review inputs for a constraint analysis, not measured performance numbers.

Run

python3 -m src.decision_matrix --scenario regulated_enterprise
python3 -m src.decision_matrix --scenario graph_heavy_orchestration --json

Validate

python3 -m pytest -o cache_dir=/tmp/agent-framework-decision-matrix-pytest-cache
python3 -m ruff check .
python3 -m compileall src tests

Scoring Model

Each framework has capability scores from 1 to 5. Each scenario provides:

• weights: importance per criterion.
• minimums: required minimum score per criterion.
• disqualify_on_minimum_failure: whether a failure should remove the framework from eligible results.

The normalized score is:

sum(capability * weight) / sum(5 * weight)

Minimum failures are reported separately so readers can understand whether a low rank is a weak fit or a hard constraint miss.