LMFlux
LMFlux is an open‑source library that helps create agentic workflows.
It supports both declarative and imperative approaches for defining LLM‑based workflows.
Features
- Declarative workflow definition
- Imperative step‑by‑step control
- Built‑in support for any OpenAI‑compatible endpoint
Quick Start
Want to get started quickly?
Install LMFlux:
pip install lmflux
Simple Python Example
from lmflux import openai_agent, tool, Session, Message
# Declare your tools
@tool
def add(a: int, b: int) -> int:
"""Add two numbers."""
return {"result": a + b}
# Create a new agent from your OAI environment variables
agent = openai_agent(
name="my‑agent",
model="openai.gpt_oss_20b",
tools=[add],
system_prompt="You are a helpful assistant. Use tools when available."
)
# Send a message to your agent
session = Session()
msg = Message("user", "What is 965 + 33264?")
resp = agent.conversate(msg, session)
print(resp.content) # → 34229
Running in a Jupyter Notebook
Below are two notebook‑friendly snippets for getting started with "agentic workflows" (Read supported architectures here)
The markdown cells (preceding each code block) explain what is happening.
1. Pre‑defined LLM tasks
# Jupyter cell: imports
from lmflux.graphs import TaskGraph, transformer_task, agentic_task
from lmflux import Session, Agent, Message, openai_agent, tool
# Jupyter cell: define a tool used by the agent
@tool
def add(a: int, b: int) -> int:
return {"result": a + b}
# Jupyter cell: create the agent
agent = openai_agent(
name="my‑agent",
model="openai.gpt_oss_20b",
tools=[add],
system_prompt="You are a helpful assistant."
)
# Jupyter cell: define tasks
@transformer_task
def set_data(session: Session):
session.set("a", 985)
session.set("b", 1265)
@agentic_task(agent) # `agent` created earlier
def add_task(agent: Agent, session: Session):
a, b = session.get("a"), session.get("b")
msg = Message("user", f"What is {a}+{b}?")
resp = agent.conversate(msg, session)
session.set("result", resp.content)
# Jupyter cell: build and run the graph, then inspect the result
G = TaskGraph()
G.connect_tasks(set_data, add_task)
G.show_mermaid() # visualises the graph
result_session = G.run()
print(result_session.context["result"]) # → 2250
graph TB
9b804ed4-64db-43f6-b69f-94411f773c98(start)
52977173-a5f2-4eb2-bb16-c7626c8e2a8b(end)
style 9b804ed4-64db-43f6-b69f-94411f773c98 fill:#ffcc00,stroke:#333,stroke-width:2px,color:#000
style 52977173-a5f2-4eb2-bb16-c7626c8e2a8b fill:#ffcc00,stroke:#333,stroke-width:2px,color:#000
28e7726e-d16d-4d1f-b499-283658471b09(set_data)
ea7a1280-3a5f-4323-a305-b697eb8e2776(add_task)
28e7726e-d16d-4d1f-b499-283658471b09 --> ea7a1280-3a5f-4323-a305-b697eb8e2776
9b804ed4-64db-43f6-b69f-94411f773c98 --> 28e7726e-d16d-4d1f-b499-283658471b09
ea7a1280-3a5f-4323-a305-b697eb8e2776 --> 52977173-a5f2-4eb2-bb16-c7626c8e2a8bTip: Use
%%captureat the top of a cell if you want to hide verbose output while still running the code.
2. Agent‑to‑Agent (Mesh) communication
from lmflux import openai_agent, tool, Session, Agent, Message
from lmflux.graphs import MeshGraph
# Mock employee database (for demo purposes)
mock_employee = {
"123456789": {"name": "Caio", "status": "ACTIVE", "department": "Engineering"},
"987654321": {"name": "Frederico", "status": "INACTIVE", "department": "Sales"},
"111111111": {"name": "Joao", "status": "ACTIVE", "department": "Accounting"},
}
@tool
def get_information(employee_id: str):
"""Retrieve employee info from a mock HR DB."""
return mock_employee.get(employee_id, {"error": "Employee not found!"})
# Create two agents
peter = openai_agent(
"peter",
"openai.gpt_oss_20b",
tools=[],
system_prompt="You are Peter, an AI agent that talks to other agents."
)
bob = openai_agent(
"bob",
"openai.gpt_oss_20b",
tools=[get_information],
system_prompt="You are Bob, an AI agent with access to employee records."
)
# Connect them in a MeshGraph
G = MeshGraph()
G.connect_agents(peter, bob, "Ask Bob about employees")
G.show_mermaid() # visual graph
graph TB
peter["peter"] -->|Can call| bob["bob"]# Run the interaction (in a notebook you can display the mermaid directly)
G.set_mermaid_render()
G.show_result()
Message(assistant):
# Run the interaction (in a notebook you can display the mermaid directly)
message = G.query_agent(peter, "What is the status of employee id 123456789?", show_progress=True)
# Message(assistant):
# content: Employee **123456789** (Caio) is **ACTIVE** and currently works in the **Engineering** department.
G.set_mermaid_render()
G.show_result()
flowchart TB
%% Global wrapper for all agents
subgraph ALL_AGENTS["All Agents"]
direction TB
subgraph subgraph_peter["peter"]
msg0["system"]
msg1["user"]
msg0 --> msg1
msg2["assistant"]
msg1 --> msg2
msg3["tool"]
msg2 --> msg3
msg4["assistant"]
msg3 --> msg4
03dd106d-cfa5-4f2a-9b12-36ae55b5a892__out(("talk to bob"))
msg2 --> 03dd106d-cfa5-4f2a-9b12-36ae55b5a892__out
end
subgraph subgraph_bob["bob"]
msg5["system"]
msg6["user"]
msg5 --> msg6
msg7["assistant"]
msg6 --> msg7
msg8["tool"]
msg7 --> msg8
msg9["assistant"]
msg8 --> msg9
03dd106d-cfa5-4f2a-9b12-36ae55b5a892__in(("talk to peter"))
03dd106d-cfa5-4f2a-9b12-36ae55b5a892__in --> msg6
msg9 --> 03dd106d-cfa5-4f2a-9b12-36ae55b5a892__in
end
end
%% Global wrapper for humans
subgraph subgraph_human["Human"]
hmsg10["User Query"]
hmsg11["User Response"]
end
class 03dd106d-cfa5-4f2a-9b12-36ae55b5a892__in interaction
class 03dd106d-cfa5-4f2a-9b12-36ae55b5a892__out interaction
03dd106d-cfa5-4f2a-9b12-36ae55b5a892__out <-.-> 03dd106d-cfa5-4f2a-9b12-36ae55b5a892__in
hmsg10 -.-> msg1
msg4 -.-> hmsg11Want more examples?
Check the Getting Started. page for additional use‑cases, deeper API explanations.