acorn

LLM agent framework with structured I/O

Build AI agents with type-safe inputs and outputs, automatic tool calling, and powerful agentic loops.

---

✨ Features

• 🎯 Structured I/O - Pydantic models for inputs and outputs
• 🤖 Agentic Loops - Multi-turn execution with tool calling
• 🛠️ Auto Tool Schemas - Generate from type hints and docstrings
• 🔄 Dynamic Tools - Add/remove tools during execution
• ✅ Parse Error Recovery - Automatic retry on validation failures
• 📊 Step Callbacks - Full control over loop behavior
• 🔌 LiteLLM Integration - Works with any LLM provider
• 🌊 Streaming Responses - Real-time output with partial structured updates
• 💾 Provider Caching - Reduce latency and cost with prompt caching
• 🛡️ Model Fallbacks - Automatic provider failover for high availability
• 🌳 Branching Workflows - Spawn sub-agents that extend parent capabilities for parallel analysis and map-reduce patterns

---

[!WARNING] Breaking Change in v0.8.0: Async-First API.

Acorn v0.8.0 makes __call__ async. If you're upgrading from v0.7.x, you need to update your code:

• Sync code: Replace agent(...) with agent.run(...)
• Async code: Add await — await agent(...)

Auto-migrate with your AI coding agent:

Paste this into your AI coding agent to update your codebase automatically:

Find all usages of acorn Module subclasses being called directly (e.g. `agent(...)`, `MyModule()(...)`).
- If the call site is in a sync function or top-level code, replace it with `agent.run(...)`.
- If the call site is in an async function, add `await` before the call: `await agent(...)`.
Do NOT change the class definitions themselves, only the call sites.

---

🚀 Quick Start

Installation

pip install acorn

Set your API key:

# For Anthropic Claude
export ANTHROPIC_API_KEY="your-key-here"

# Or for OpenAI
export OPENAI_API_KEY="your-key-here"

# Or any other LiteLLM-supported provider

---

Single-Turn Example

import asyncio
from pydantic import BaseModel, Field
from acorn import Module

class Input(BaseModel):
    text: str = Field(description="The text to summarize")
    max_words: int = Field(default=100, description="Maximum words in summary")

class Output(BaseModel):
    summary: str = Field(description="The concise summary")
    word_count: int = Field(description="Number of words in summary")

class Summarizer(Module):
    """Summarize text concisely."""

    initial_input = Input
    final_output = Output
    model = "anthropic/claude-sonnet-4-5-20250514"

# Use it (async)
async def main():
    summarizer = Summarizer()
    result = await summarizer(
        text="Long article text here...",
        max_words=50
    )
    print(result.summary)
    print(f"Words: {result.word_count}")

asyncio.run(main())

For synchronous code, use .run():

# Use it (sync)
summarizer = Summarizer()
result = summarizer.run(
    text="Long article text here...",
    max_words=50
)
print(result.summary)
print(f"Words: {result.word_count}")

Multi-Turn Agentic Loop

import asyncio
from pydantic import BaseModel, Field
from acorn import Module, tool

class Input(BaseModel):
    topic: str = Field(description="Research topic")
    depth: str = Field(default="shallow", description="Research depth")

class Output(BaseModel):
    findings: str = Field(description="Summary of findings")
    sources: list[str] = Field(description="Sources consulted")

class ResearchAgent(Module):
    """Research assistant with tools."""

    initial_input = Input
    max_steps = 5  # Enable agentic loop
    final_output = Output
    model = "anthropic/claude-sonnet-4-5-20250514"

    @tool
    def search(self, query: str) -> list:
        """Search for information."""
        # Your search implementation
        return ["result1", "result2"]

    @tool
    def analyze(self, data: str) -> str:
        """Analyze collected data."""
        # Your analysis implementation
        return f"Analysis: {data}"

    def on_step(self, step):
        """Called after each step."""
        print(f"Step {step.counter}")

        # Early termination if done
        if len(step.tool_results) >= 3:
            step.finish(
                findings="Sufficient data collected",
                sources=["source1", "source2"]
            )

        return step

# Use it (async)
async def main():
    agent = ResearchAgent()
    result = await agent(topic="Large Language Models", depth="shallow")

asyncio.run(main())

For synchronous code:

# Use it (sync)
agent = ResearchAgent()
result = agent.run(topic="Large Language Models", depth="shallow")

---

📖 Documentation

askmanu.github.io/acorn

• Getting Started - Installation and first steps
• Module Reference - Complete Module API documentation
• Branching - Sub-agents and parallel processing

---

📚 Core Concepts

Module

Base class for LLM agents. Configure with:

• model - LLM to use (required - no default)
• temperature - Sampling temperature
• max_tokens - Maximum tokens to generate
• max_steps - Max agentic loop iterations (None = single-turn)
• initial_input - Pydantic model for input schema
• final_output - Pydantic model for output schema
• tools - List of available tools
• cache - Enable provider-level prompt caching
• model_fallbacks - List of fallback models for automatic failover

Tools

Functions the LLM can call:

@tool
def search(query: str, limit: int = 10) -> list:
    """Search for information.

    Args:
        query: The search query
        limit: Maximum results to return
    """
    return search_api(query, limit)

Schema is automatically generated from type hints and docstring.

Tools can be async:

@tool
async def fetch_data(url: str) -> dict:
    """Fetch data from a URL.

    Args:
        url: The URL to fetch from
    """
    async with httpx.AsyncClient() as client:
        response = await client.get(url)
        return response.json()

When the LLM calls multiple tools in one step, they execute concurrently via asyncio.gather() for better performance.

Step Callback

Control agentic loop execution:

def on_step(self, step):
    # Access step info
    print(f"Step {step.counter}")
    print(f"Tools called: {[tc.name for tc in step.tool_calls]}")

    # Dynamic tool management
    step.add_tool(new_tool)
    step.remove_tool("old_tool")

    # Early termination
    if condition:
        step.finish(result="Early exit")

    return step

---

🎯 Examples

Try them live on the Gradio app or browse the source in examples/:

---

🧪 Testing

# Run all tests
pytest

# With coverage
pytest --cov=acorn

# Specific test file
pytest tests/test_agentic_loop.py -v

Current status: 201 tests passing, 85% coverage

---

🛣️ Roadmap

✅ Completed

• Single-turn execution
• Multi-turn agentic loops
• Tool calling with auto-schema generation
• Parse error recovery
• Dynamic tool management
• Step callbacks
• Streaming responses with partial structured output
• Forced termination strategies
• Provider caching
• Model fallbacks
• Branching workflows
• Async-first API with sync wrapper

📋 Planned

• More docs
• Integration examples with different providers (vector DBs, observability tools, etc.)

---

🤝 Contributing

Contributions welcome! Please:

1. Check open issues for areas to help
2. Write tests for new features (maintain >80% coverage)
3. Update documentation
4. Add examples for new features

---

🙏 Acknowledgments

Thanks to @rosenbrockc for donating the acorn pip package name.

---

📄 License

MIT License