🚀 Overview

Universal Agent File (UAF) is a standardized binary format for packaging and distributing AI agents.

It solves a core problem in the agent ecosystem:

❌ Agents are fragmented, framework-locked, and not portable ✅ UAF makes agents portable, reproducible, and shareable

A .uaf file is a compressed, verifiable artifact that bundles agent logic, dependencies, metadata, tools, and optional state.

This enables: Write once → run anywhere → resume anytime

✨ Features

• 📦 Portable Binary Format - Package complete agents into a single .uaf file.
• 🧠 Stateful Agents - Persist and transfer agent memory across environments.
• 🛡️ Strict Validation - Enforced schema (agent.yaml) ensures reliability.
• 🔍 Inspect Without Running - View metadata, tools, and config without execution.
• 🔌 Framework Agnostic - Works across LangChain, LangGraph, CrewAI, Google ADK.
• ⚡ Runtime Loader - Dynamically load and execute agents from .uaf.

🛠️ Installation

Install from PyPI (Python 3.9+):

pip install uaf_cli

Or from source:

git clone https://github.com/vaibhavhaswani/uaf-cli.git
cd uaf-cli
pip install .

⚡ Quick Start

1️⃣ Project Structure

Create an agent folder with the following structure:

my-agent/
├── agent.py          # Framework logic (LangChain, CrewAI, etc.)
├── agent.yaml        # Manifest metadata
├── requirements.txt  # Dependencies
├── agent.state       # (Optional) Initial state
└── uaf_setup.yaml    # Compiler build instructions

2️⃣ Configure Build (uaf_setup.yaml)

output: my-agent.uaf

files:
  agent.py: ./agent.py
  agent.yaml: ./agent.yaml
  requirements.txt: ./requirements.txt
  agent.state: ./agent.state

3️⃣ Compile

uaf compile -f uaf_setup.yaml

✅ Output: my-agent.uaf

💻 CLI

Compile, package, and execute agents with simple commands:

uaf init        # Scaffold absolute path project
uaf compile     # Compile into .uaf artifact
uaf validate    # Validate schema compliance
uaf inspect     # View deep metadata
uaf run         # Run CLI agent turn loop
uaf update      # Inject incremental builds

Full technical guide → docs/usage_guide.md

🔗 Loading Agents

The UAFLoader dynamically routes directly into execution setups powered by target frame pipelines.

from uaf_compiler.loader import UAFLoader

# 1. Initialize Loader
uaf_loader = UAFLoader("my-agent.uaf")

# 2. Dynamic Router Injection
# Load LangChain/LangGraph
agent_app = uaf_loader.load(llm=my_llm)

# Load CrewAI
agent_crew = uaf_loader.load(llm_model="ollama/gemma3", base_url="http://...")

# Load Google ADK
agent_runner = uaf_loader.load(llm_model="ollama/gemma3", base_url="http://...")

Full Loading & Integration workflows → docs/integration.md

📜 Protocol Specification

Complete binary spec design detailing underlying tarball structure, manifest validation rules, and lifecycle bindings.

See frameworks & schema docs → docs/frameworks.md

🏗️ Development & Testing

Standard installation and developer workflows:

pip install -e .

Run End-to-End Tests

To verify loaders and correct tool-use turn loops across all SDK adaptors (LangChain, CrewAI, ADK, Comet) over a local Ollama endpoint:

python testing/test_dummy_agents.py

AgentComet natively loads and executes UAF agents using its SDK-level hooks without wrapping logic adapters.

from agentcomet import load_agent

# 1. Load UAF directly 
agent = load_agent("my-agent.uaf")

# 2. Run agent natively
response = agent.run("Calculate current stock performance")
print("Response:", response)

# 3. Export / Save state
agent.export("my_assistant.uaf")