voluntas 0.1.0

A BDI (Belief-Desire-Intention) agent framework built on Pydantic AI

Voluntas

Voluntas is a BDI (Belief-Desire-Intention) agent framework built on top of Pydantic AI. It provides structured beliefs, desires, intentions, adaptive planning, execution, reconsideration, usage tracking, and optional human-in-the-loop intervention.

Installation

pip install voluntas

Or with uv:

uv add voluntas

The distribution name is voluntas and the Python import is also voluntas.

Quick start

import asyncio

from pydantic_ai.models.test import TestModel
from voluntas import BDI


async def main() -> None:
    agent = BDI(
        model=TestModel(),
        desires=["Prepare a concise project status report"],
        intentions=["Inspect the available project information"],
    )

    await agent.bdi_cycle()


asyncio.run(main())

For production use, replace TestModel with a model supported by Pydantic AI and install any provider-specific dependencies required by that model.

Public API

The main agent and commonly used schemas are available from the package root:

from voluntas import (
    BDI,
    BDIUsageTracker,
    Belief,
    BeliefSet,
    Desire,
    DesireStatus,
    Intention,
    Plan,
)

The complete schema surface is available from voluntas.schemas:

from voluntas.schemas import (
    BeliefExtractionResult,
    HighLevelIntentionList,
    PlanManipulationDirective,
    ReconsiderResult,
)

BDI lifecycle

Each cycle coordinates the following stages:

1. Update beliefs from the current context and action outcomes.
2. Deliberate over pending desires and their priorities.
3. Generate a high-level intention when no active intention exists.
4. Execute one intention step using Pydantic AI tools and toolsets.
5. Reconsider the remaining plan after failed or changed work.

The framework supports MCP servers through the Pydantic AI integration passed to BDI, as well as structured logs and aggregate usage tracking through BDIUsageTracker.

Human-in-the-loop

Set enable_human_in_the_loop=True to allow failures to be presented to a human for guidance. The guidance is interpreted into structured actions such as retrying, modifying, replacing, inserting, skipping, or aborting plan steps.

Development

Clone the repository and install development dependencies with uv:

uv sync --group dev
uv run pytest

The repository also contains SBench and benchmark runners for research and experiments. The runners use a local LiteLLM proxy that exposes an OpenAI-compatible API; they are development applications and are not part of the published voluntas package.

Set the proxy connection before running the local examples:

export LITELLM_BASE_URL=http://localhost:4000
export LITELLM_API_KEY=sk-1234
export LITELLM_MODEL=gpt-5.3-codex

The value of LITELLM_MODEL must match a model alias configured in the proxy.

License

Voluntas is released under the MIT license.