Hanzo Agent SDK

hanzo-agent runs agents in Python. An agent is a model reached over an OpenAI-compatible /v1 API, its instructions, and the Python functions it may call as tools. It installs as hanzo-agent and imports as agents.

Install

Python 3.9 or newer.

pip install hanzo-agent

First agent

from openai import AsyncOpenAI
from agents import Agent, Runner, set_default_openai_api, set_default_openai_client

set_default_openai_client(AsyncOpenAI(base_url="http://localhost:11434/v1", api_key="ollama"))
set_default_openai_api("chat_completions")

agent = Agent(name="Assistant", instructions="You are a helpful assistant.", model="qwen3:0.6b")
result = Runner.run_sync(agent, "Write a haiku about recursion in programming.")
print(result.final_output)

One run printed:

Recursive function deferring calls,
A loop unrolled with grace,
Ends with a loop.

The two calls before the agent decide where it runs:

• set_default_openai_client names the endpoint and the key. Above it is a local server on port 11434, Ollama with qwen3:0.6b pulled. For the hosted API pass base_url="https://api.hanzo.ai/v1", a Hanzo API key as api_key, and a model id from https://api.hanzo.ai/v1/models.
• set_default_openai_api("chat_completions") sends POST /v1/chat/completions. Without it the SDK sends POST /v1/responses instead.

Tools

A decorated function becomes a tool the model can call:

from agents import function_tool

@function_tool
def get_weather(city: str) -> str:
    return f"The weather in {city} is sunny."

agent = Agent(
    name="Weather",
    instructions="Answer with the get_weather tool.",
    tools=[get_weather],
    model="qwen3:0.6b",
)
print(Runner.run_sync(agent, "What is the weather in Tokyo?").final_output)

With the setup above it printed The weather in Tokyo is sunny.

Errors and retries

Model calls go through the openai client and follow its rules. A 401, 402 or 403 raises on the first answer (AuthenticationError, APIStatusError, PermissionDeniedError). A 429 or 5xx is retried twice and then raises; AsyncOpenAI(..., max_retries=0) sends it once. A 200 whose body is an error object raises openai.APIError carrying the server's message, as the client does for an error inside a stream. A <think>…</think> block at the start of a reply is dropped from final_output and from the history the next turn sends; result.raw_responses keeps it.

Tracing

Runs create traces and spans and send them nowhere. To export them, add a processor:

from agents import add_trace_processor
from agents.tracing.processors import BackendSpanExporter, BatchTraceProcessor

exporter = BackendSpanExporter(endpoint="https://collector.example/ingest", api_key="...")
add_trace_processor(BatchTraceProcessor(exporter))

The exporter posts batches of JSON to that endpoint with the key as a bearer token. set_tracing_disabled(True) or OPENAI_AGENTS_DISABLE_TRACING=1 stops creating them.

Extras

pip install "hanzo-agent[web3]"   # agents.extensions.web3: AgentWallet, Web3Wallet, MpcClient
pip install "hanzo-agent[tee]"    # agents.extensions.tee: ConfidentialAgent, TEEProvider

agents.network routes a request among several agents and agents.orchestration runs workflows over registered agents; both are in the base package.

This repository

The root is also a Go module, github.com/hanzoai/agent, and sdk/ holds Python, TypeScript and Go clients published under other names. This README covers hanzo-agent only.

Development

make sync    # uv sync --all-extras --all-packages --group dev
make tests   # uv run pytest