Reagentic Framework

This is a home-made openai-agents based framework for simple agent coding. I am writing it for myself, so as they say, no guarantees.

Installation

pip install -r requirements.txt
pip install -e .

Providers

Now the is only one provider - Openrouter. It has automatic refiller, so, in theory, there will be always actual models.

Settings

The framework rely on env var for providers key. It can be placed in environment, local .env (in project), .env near .venv

Usage

Simple hello world

from agents import Agent, Runner
import reagentic.providers.openrouter as openrouter   

provider = openrouter.OpenrouterProvider(openrouter.DEEPSEEK_CHAT_V3_0324)
agent = Agent(name="Assistant", instructions="You are a helpful assistant", model=provider.get_openai_model())

result = Runner.run_sync(agent, "Write a haiku about recursion in programming.")
print(result.final_output)

or with auto choose of model:

from agents import Agent, Runner
from reagentic.providers import openrouter

provider = openrouter.auto()
print(f"Will use {provider.model} model")
agent = Agent(name="Assistant", instructions="You are a helpful assistant", model=provider.get_openai_model())

result = Runner.run_sync(agent, "Write a haiku about recursion in programming.")
print(result.final_output)

Call some predefined script:

refill openrouter models

python3 -m reagentic --script openrouter_refiller

choose some model (minimalistic, but free)

python3 -m reagentic --script model_chooser --provider openrouter --request "latest deepseek model"

Example output:
Chosen Model: DEEPSEEK_CHAT

choose some model. More smart, but paid

python -m reagentic --script model_chooser --provider openrouter --request "latest deepseek model" --paid true

Example output:

Chosen Model: deepseek/deepseek-chat-v3-0324
Reason: DeepSeek Chat V3 0324 is the latest iteration of DeepSeek models, offering a massive 685B-parameter MoE 
architecture with outstanding instruction following and coding capabilities. This model succeeds the previous DeepSeek V3 and delivers strong performance across a variety of tasks, making it ideal for users seeking cutting-edge open-source DeepSeek models.