build-agent

Building Agents with orxhestra

All agents extend BaseAgent and implement astream(input, *, ctx) returning AsyncIterator[Event].

LlmAgent — Standard tool-calling agent

from orxhestra import LlmAgent
from langchain_core.tools import tool
from langchain_openai import ChatOpenAI

@tool
async def search(query: str) -> str:
    """Search the web."""
    return f"Results for: {query}"

agent = LlmAgent(
    name="assistant",
    model=ChatOpenAI(model="gpt-5.4"),
    tools=[search],
    instructions="You are a helpful assistant.",
    max_iterations=10,
)

# Async streaming
async for event in agent.astream("Hello"):
    if event.is_final_response():
        print(event.text)

# Sync convenience
result = agent.invoke("Hello")
print(result.text)

Key parameters

Parameter	Type	Description
name	str	Unique agent name
model	BaseChatModel	Any LangChain chat model
tools	list[BaseTool]	Tools available to the agent
instructions	str | Callable	System prompt (static or dynamic)
planner	BasePlanner	Optional planning strategy
output_schema	type	Optional Pydantic model for structured output
max_iterations	int	Max tool-call loop iterations (default: 10)

Dynamic instructions

async def dynamic_instructions(ctx):
    return f"You are helping user in session {ctx.session_id}."

agent = LlmAgent(
    name="dynamic",
    model=model,
    instructions=dynamic_instructions,
)

ReActAgent — Structured reasoning

Uses with_structured_output() to enforce a typed ReActStep at every iteration. Extends LlmAgent so it inherits instructions, planners, skills, and callbacks.

from orxhestra import ReActAgent

agent = ReActAgent(
    name="reasoner",
    model=ChatOpenAI(model="gpt-5.4"),
    tools=[search],
    instructions="Think step by step.",  # appended to ReAct prompt
    max_iterations=10,
)

Running with sessions (Runner)