Goal
By the end of this page, you will:
- Wrap Python functions as
Steps
- Compose steps into a workflow
- Add an LLM step
- Add tool steps
- Run a prebuilt ReAct agent
Minimal workflow (Step → compose)
Create a step from a function, then run it in a simple sequence.
from coevolved.base import run_sequence, step
@step(name="normalize")
def normalize(state: dict) -> dict:
text = (state.get("text") or "").strip()
return {**state, "text": text}
@step(name="classify")
def classify(state: dict) -> dict:
category = "empty" if not state["text"] else "non-empty"
return {**state, "category": category}
workflow = [normalize, classify]
final_state = run_sequence(workflow, {"text": " hello "})
Add an LLM step
Use llm_step(...) to turn “prompt building + provider call” into a Step.
from openai import OpenAI
from coevolved.core import LLMConfig, OpenAIProvider, llm_step
provider = OpenAIProvider(OpenAI())
config = LLMConfig(model="gpt-4o-mini", temperature=0)
def build_prompt(state: dict) -> str:
return f"Rewrite this in one short sentence: {state['text']}"
rewrite = llm_step(
prompt_builder=build_prompt,
provider=provider,
config=config,
name="rewrite",
)
kind="tool" and optional argument schemas for function calling.
from pydantic import BaseModel
from coevolved.core import tool_step
class AddArgs(BaseModel):
a: int
b: int
def add_tool(state: dict) -> int:
args = state["tool_args"]
return int(args["a"]) + int(args["b"])
add = tool_step(
add_tool,
name="add",
tool_schema=AddArgs,
result_key="tool_result",
)
Run a prebuilt agent (ReAct)
The prebuilt ReAct agent alternates between:
- an LLM “planner” step
- executing a tool when the LLM requests it
from coevolved.core import tool_specs_from_dict
from coevolved.prebuilt import react_agent
tools = {"add": add}
planner_config = config.model_copy(update={"tools": tool_specs_from_dict(tools)})
planner = llm_step(
prompt_builder=lambda s: s["messages"],
provider=provider,
config=planner_config,
name="planner",
)
agent = react_agent(planner=planner, tools=tools, max_steps=5)
result = agent(
{
"messages": [
{"role": "user", "content": "What is 2 + 3? Use the add tool."}
]
}
)
ReAct expects messages to be present if you want the loop to accumulate conversation context.
Next steps
