Vibe Agents

VibeAgent allows you to build and execute multi-modal flows using natural language descriptions. Instead of manually defining tasks and dependencies, you describe your intent, and Vibe Agent handles the orchestration using LLMs.

Your intent is compiled into execution graph, each node represents an agent or tool action, and each edge encodes dependencies and data flow.

info

VibeAgent is beta supported starting from version 1.4.0 as we work toward AGI where agents generate agents.

How it works

1. Planner: A high-level LLM (OpenAI, Gemini, or Anthropic) analyzes your description.
2. Spec Generation: It generates a structured FlowSpec JSON containing tasks, agents, and routing.
3. Flow Building: VibeFlow converts the spec into a real Flow object.
4. Execution: You start the flow with your initial input.

Simple Example

Build a text-based joke generator using a natural language "vibe":

import asyncio
import os
from intelli.flow import VibeAgent

async def main():
    # 1. Setup the planner (requires an API key)
    vf = VibeAgent(
        planner_provider="gemini",
        planner_api_key=os.getenv("GEMINI_API_KEY"),
        planner_model="gemini-2.0-flash"
    )
    
    # 2. Describe the flow you want
    description = "Create a 1-step flow that returns a funny joke about AI agents."
    
    # 3. Build the flow
    flow = await vf.build(description)
    
    # 4. Execute
    results = await flow.start(initial_input="Tell me a joke")
    
    for name, data in results.items():
        print(f"Result: {data['output']}")

if __name__ == "__main__":
    asyncio.run(main())

Multi-Modal Vibe

VibeAgent can orchestrate different types of agents (text, image, audio) in a single request:

description = (
    "1. Generate a speech audio for 'Intelli is awesome' using tts-1. "
    "2. Transcribe that audio back to text using whisper-1."
)

flow = await vf.build(description)
results = await flow.start()

Auto-Saving Outputs

VibeAgent can automatically save generated images and audio to a specific directory:

flow = await vf.build("Generate an image of a cyber cat and save it to ./outputs")
flow.auto_save_outputs = True
flow.output_dir = "./outputs"

await flow.start()

Preferred Models

You can specify preferred model details as strings when initializing VibeAgent. This guides the planner to use specific versions of AI models for text, image, speech, or recognition tasks.

vf = VibeAgent(
    planner_provider="gemini",
    planner_api_key=os.getenv("GEMINI_API_KEY"),
    # Specify specific model versions as descriptive strings
    text_model="openai gpt-5.2-mini",
    image_model="gemini gemini-3-pro-image-preview",
    speech_model="openai tts-1",
    recognition_model="openai whisper-1"
)

Supported preference parameters:

• text_model
• image_model
• speech_model
• recognition_model

Key Features

• Environment Variables: Use ${ENV:VARIABLE_NAME} in your prompts, and VibeAgent will automatically resolve them from your .env file (secrets stay hidden).
• Save & Load: Use save_bundle(save_dir) to export the generated flow spec and a graph image visualization for future use.
• Edit Mode: Call vf.edit(spec_path, "add a translation step") to modify an existing flow using natural language.

Detailed Example

For a working end-to-end example (search → analyze → image generation + auto-saving), see:
sample/vibeagent_research_factory.py