Vibe Features

Vibe Agent uses a high-level planner to architect specialized multi-agent workflows. This page covers the technical design behind graph generation, tool integration, and flow lifecycle management.

Initialization Parameters

The VibeAgent accepts the following configuration parameters:

Parameter	Type	Description
planner_provider	str	LLM provider for the planner (openai, anthropic, gemini, vllm, llamacpp).
planner_api_key	str	API key for cloud planner providers. Optional for vllm or llamacpp.
planner_model	str	Specific model version (e.g., gpt-4o, gemini-2.0-flash).
text_model	str	(Optional) Preferred model string for text tasks.
image_model	str	(Optional) Preferred model string for image tasks.
speech_model	str	(Optional) Preferred model string for speech tasks.
recognition_model	str	(Optional) Preferred model string for transcription tasks.
processors	dict	Mapping of strings to post-processing functions.
context_files	list	Custom list of code files to provide to the planner for context.

Quick Example

from intelli.flow import VibeAgent

# 1. Initialize the architect
va = VibeAgent(planner_provider="openai", planner_api_key="your_key")

# 2. Build the graph from a vibe
flow = await va.build("Research AI trends and generate a summary report")

# 3. Execute the flow
results = await flow.start()

The Generated Blueprint

The blueprint is the bridge between natural language and the Flow execution engine.

Task Definition

Each task in the blueprint follows this schema:

• name: Unique identifier for the task node.
• desc: The instruction to be executed by the agent.
• agent: AgentSpec containing agent_type, provider, and mission.
• model_params: Dictionary of model-specific parameters (e.g., temperature, max_tokens).
• post_process: Key of a function defined in the VibeAgent.processors map.

Global Configuration

• max_workers: Integer defining the maximum parallel task execution.
• output_memory_map: Maps task outputs to shared memory keys for downstream access.
• map_paths: Defines the dependency edges between tasks (DAG structure).

Advanced Methods

Flow Persistence

Use save_bundle to export the generated blueprint for production use without re-running the planner.

# Exports flow_spec.json and vibeflow_graph.png
va.save_bundle(save_dir="./prod_flow", spec=va.last_spec, flow=flow)

Direct Blueprint Execution

Reconstruct a flow directly from a saved blueprint to save latency and costs.

spec = va.load_spec("./prod_flow/flow_spec.json")
flow = va.build_from_spec(spec)

Blueprint Modification

Modify existing architectures using natural language instructions.

# existing_spec is loaded from a path or va.last_spec
updated_flow = await va.edit(spec_path="./path/to/spec.json", instruction="Add a validation step")

Runtime Features

Environment Variable Injection

Vibe Agent supports secret resolution at runtime using the ${ENV:VAR_NAME} syntax. This ensures that API keys or sensitive endpoints are not persisted in the flow_spec.json file.

Automatic Output Management

Configure the execution engine to handle file persistence automatically:

• flow.auto_save_outputs = True: Automatically writes images and audio files to disk.
• flow.output_dir = "./path": Sets the target directory for all generated assets.

Offline / Local Models (vLLM + llama.cpp)

VibeAgent supports local inference for the Architect (planner) or individual tasks when you provide the connection details.

• vLLM: provide baseUrl (e.g. ${ENV:VLLM_BASE_URL}) and a model name.
• llama.cpp: provide model_path (e.g. ${ENV:LLAMACPP_MODEL_PATH}).

import os
from intelli.flow import VibeAgent

va = VibeAgent(
    planner_provider="vllm",
    planner_model="deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
    planner_options={"baseUrl": os.getenv("VLLM_BASE_URL")},
)

Example prompt to generate offline agent

This prompt tells the Architect (planner) to design the execution graph and generate with offline text agent.
When you call await va.build(vibe), VibeAgent uses the planner to produce the full spec (tasks, providers, routing) and then builds a runnable Flow.

vibe = (
    "Create a 2-step flow. "
    "Step 1 (research): Use OpenAI to extract 8 key facts about solid-state batteries. "
    "Step 2 (offline_summary): Use vLLM at ${ENV:VLLM_BASE_URL} with model "
    "'deepseek-ai/DeepSeek-R1-Distill-Llama-8B' to summarize the facts into 5 bullets."
)

flow = await va.build(vibe)