AdalFlow

Why AdalFlow

1. 100% Open-source Agents SDK: Lightweight and requires no additional API to setup Human-in-the-Loop and Tracing Functionalities.
2. Say goodbye to manual prompting: AdalFlow provides a unified auto-differentiative framework for both zero-shot optimization and few-shot prompt optimization. Our research, LLM-AutoDiff and Learn-to-Reason Few-shot In Context Learning, achieve the highest accuracy among all auto-prompt optimization libraries.
3. Switch your LLM app to any model via a config: AdalFlow provides Model-agnostic building blocks for LLM task pipelines, ranging from RAG, Agents to classical NLP tasks.

View Documentation

Quick Start

Install AdalFlow with pip:

pip install adalflow

Hello World Agent Example

from adalflow import Agent, Runner
from adalflow.components.model_client.openai_client import OpenAIClient
from adalflow.core.types import (
    ToolCallActivityRunItem, 
    RunItemStreamEvent,
    ToolCallRunItem,
    ToolOutputRunItem,
    FinalOutputItem
)
import asyncio

# Define tools
def calculator(expression: str) -> str:
    """Evaluate a mathematical expression."""
    try:
        result = eval(expression)
        return f"The result of {expression} is {result}"
    except Exception as e:
        return f"Error: {e}"

async def web_search(query: str="what is the weather in SF today?") -> str:
    """Web search on query."""
    await asyncio.sleep(0.5)
    return "San Francisco will be mostly cloudy today with some afternoon sun, reaching about 67 °F (20 °C)."

def counter(limit: int):
    """A counter that counts up to a limit."""
    final_output = []
    for i in range(1, limit + 1):
        stream_item = f"Count: {i}/{limit}"
        final_output.append(stream_item)
        yield ToolCallActivityRunItem(data=stream_item)
    yield final_output

# Create agent with tools
agent = Agent(
    name="MyAgent",
    tools=[calculator, web_search, counter],
    model_client=OpenAIClient(),
    model_kwargs={"model": "gpt-4o", "temperature": 0.3},
    max_steps=5
)

runner = Runner(agent=agent)

1. Synchronous Call Mode

# Sync call - returns RunnerResult with complete execution history
result = runner.call(
    prompt_kwargs={"input_str": "Calculate 15 * 7 + 23 and count to 5"}
)

print(result.answer)
# Output: The result of 15 * 7 + 23 is 128. The counter counted up to 5: 1, 2, 3, 4, 5.

# Access step history
for step in result.step_history:
    print(f"Step {step.step}: {step.function.name} -> {step.observation}")
# Output:
# Step 0: calculator -> The result of 15 * 7 + 23 is 128
# Step 1: counter -> ['Count: 1/5', 'Count: 2/5', 'Count: 3/5', 'Count: 4/5', 'Count: 5/5']

2. Asynchronous Call Mode

# Async call - similar output structure to sync call
result = await runner.acall(
    prompt_kwargs={"input_str": "What's the weather in SF and calculate 42 * 3"}
)

print(result.answer)
# Output: San Francisco will be mostly cloudy today with some afternoon sun, reaching about 67 °F (20 °C). 
#         The result of 42 * 3 is 126.

3. Async Streaming Mode

# Async streaming - real-time event processing
streaming_result = runner.astream(
    prompt_kwargs={"input_str": "Calculate 100 + 50 and count to 3"},
)

# Process streaming events in real-time
async for event in streaming_result.stream_events():
    if isinstance(event, RunItemStreamEvent):
        if isinstance(event.item, ToolCallRunItem):
            print(f"🔧 Calling: {event.item.data.name}")
        elif isinstance(event.item, ToolCallActivityRunItem):
            print(f"📝 Activity: {event.item.data}")
        elif isinstance(event.item, ToolOutputRunItem):
            print(f"✅ Output: {event.item.data.output}")
        elif isinstance(event.item, FinalOutputItem):
            print(f"🎯 Final: {event.item.data.answer}")

# Output:
# 🔧 Calling: calculator
# ✅ Output: The result of 100 + 50 is 150
# 🔧 Calling: counter
# 📝 Activity: Count: 1/3
# 📝 Activity: Count: 2/3
# 📝 Activity: Count: 3/3
# ✅ Output: ['Count: 1/3', 'Count: 2/3', 'Count: 3/3']
# 🎯 Final: The result of 100 + 50 is 150. Counted to 3 successfully.

Set your OPENAI_API_KEY environment variable to run these examples.

Try the full Agent tutorial in Colab:

View Quickstart: Learn How AdalFlow optimizes LM workflows end-to-end in 15 mins.

Go to Documentation for tracing, human-in-the-loop, and more.

Research

[Sep 2025] LAD-VF: LLM-Automatic Differentiation Enables Fine-Tuning-Free Robot Planning from Formal Methods Feedback

• Fine-tuning-free robot planning using LLM auto-differentiation
• Integration of formal methods feedback for robot control

[Jan 2025] Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting

• LLM Applications as auto-differentiation graphs
• Token-efficient and better performance than DsPy

[Dec 2025] Scaling Textual Gradients via Sampling-Based Momentum

• Stable, scalable prompt optimization using momentum-weighted textual gradient
• Gumbel-Top-k sampling improves exploration and integrates seamlessly with TextGrad, DSPy-COPRO, and AdalFlow

Auto-Prompt Optimization Ecosystem

AdalFlow is part of a growing ecosystem of libraries that automatically optimize LLM prompts and workflows. Here's how the landscape looks:

Library	Approach	Key Idea
AdalFlow	PyTorch-style auto-differentiation	LLM workflows as auto-diff graphs; unified textual gradient descent + few-shot bootstrap optimization in one training loop
DSPy	Declarative programming	Write compositional Python code instead of prompts; compiler optimizes prompts and weights automatically
Agent Lightning	Framework-agnostic agent trainer	Turn any agent (LangChain, OpenAI SDK, AutoGen, etc.) into an optimizable entity with minimal code changes; supports RL, auto-prompt optimization, and supervised fine-tuning
TextGrad	Textual gradient descent	Automatic differentiation via text; uses LLM feedback as gradients to optimize prompts, code, and solutions

Where AdalFlow fits: AdalFlow draws inspiration from all of the above (see Acknowledgements) and unifies them into a single PyTorch-like framework. You get textual gradients (à la TextGrad), few-shot bootstrap (à la DSPy), and instruction history — all composable within Parameter, Generator, AdalComponent, and Trainer.

Collaborations

We work closely with the VITA Group at University of Texas at Austin, under the leadership of Dr. Atlas Wang and in collaboration with Dr. Junyuan Hong, who provides valuable support in driving project initiatives.

For collaboration, contact Li Yin.

Hiring

We are looking for a Dev Rel to help us build the community and support our users. If you are interested, please contact Li Yin.

Documentation

AdalFlow full documentation available at adalflow.sylph.ai:

AdalFlow: A Tribute to Ada Lovelace

AdalFlow is named in honor of Ada Lovelace, the pioneering female mathematician who first recognized that machines could go beyond mere calculations. As a team led by a female founder, we aim to inspire more women to pursue careers in AI.

Community & Contributors

The AdalFlow is a community-driven project, and we welcome everyone to join us in building the future of LLM applications.

Join our Discord community to ask questions, share your projects, and get updates on AdalFlow.

To contribute, please read our Contributor Guide.

Contributors

Acknowledgements

Many existing works greatly inspired AdalFlow library! Here is a non-exhaustive list:

• 📚 PyTorch for design philosophy and design pattern of Component, Parameter, Sequential.
• 📚 Micrograd: A tiny autograd engine for our auto-differentiative architecture.
• 📚 Text-Grad for the Textual Gradient Descent text optimizer.
• 📚 DSPy for inspiring the __{input/output}__fields in our DataClass and the bootstrap few-shot optimizer.
• 📚 OPRO for adding past text instructions along with its accuracy in the text optimizer.
• 📚 PyTorch Lightning for the AdalComponent and Trainer.