Model

Models are the LLM backends that power agent reasoning. Agiwo keeps them behind one streaming-first abstraction.

Core contract

Every provider implements:

async def arun_stream(
    self,
    messages: list[dict],
    tools: list[dict] | None = None,
) -> AsyncIterator[StreamChunk]:
    ...

StreamChunk is the normalized delta format that the rest of the runtime consumes.

Provider choices

Public providers include:

OpenAIModel
OpenAIResponsesModel
AnthropicModel
DeepseekModel
NvidiaModel
BedrockAnthropicModel

Compatible endpoints are supported through the model factory.

Recommended construction path

Use the factory when the model comes from config or user input:

from agiwo.llm import create_model_from_dict

model = create_model_from_dict(
    provider="openai-compatible",
    model_name="my-model",
    params={
        "base_url": "https://api.example.com/v1",
        "api_key_env_name": "MY_API_KEY",
    },
)

Use direct classes when the model is hard-coded in application code:

from agiwo.llm import OpenAIModel

model = OpenAIModel(name="gpt-5.4")

Why the abstraction matters

The rest of Agiwo depends on one normalized streaming interface instead of provider-specific response types. That is what makes agent execution, tool calling, scheduler orchestration, and tracing work across providers.