fastmind 0.2.6

A lightweight, event-driven multi-agent framework for embodied AI systems

FastMind 🧠

A lightweight, event-driven framework for building embodied AI agents with dual-loop architecture (LLM + VLA).

Features

• Dual-Loop Architecture: LLM slow loop (planning, reasoning) + VLA fast loop (real-time control) run concurrently in one session
• Signal & Event: Two parallel communication primitives — Event for discrete messages, Signal for continuous high-frequency data
• FastAPI-like Decorators: Familiar @app.agent, @app.tool, @app.vla, @app.vla_action, @app.signal syntax
• State Graph: Build agent workflows like flowcharts with nodes, edges, and conditional routing
• Event-Driven: Asyncio-based, zero polling, high-performance async execution
• Human-in-the-Loop: Interrupt and resume sessions for human approval
• Perception Loops: Native support for sensors, timers, and external triggers
• Action Channel Routing: VLA outputs map to multiple action executors via channel names (N:M)
• Session Isolation: Multi-user support with isolated state per session
• Lightweight: ~8000 lines core, no big dependencies

Installation

pip install fastmind

Quick Start

LLM Agent

from fastmind import FastMind, Graph, Event
from fastmind.contrib import FastMindAPI

app = FastMind()

@app.agent(name="chat_agent")
async def chat_agent(state: dict, event: Event) -> dict:
    state.setdefault("messages", [])
    state["messages"].append({"role": "user", "content": event.payload.get("text", "")})
    state["messages"].append({"role": "assistant", "content": "Hello!"})
    return state

graph = Graph()
graph.add_node("chat", chat_agent)
graph.set_entry_point("chat")
app.register_graph("main", graph)

async def main():
    api = FastMindAPI(app)
    await api.start()
    await api.push_event("user_001", Event("user.message", {"text": "Hello!"}, "user_001"))
    await api.stop()

import asyncio
asyncio.run(main())

LLM Agent with Tool Calling (ReAct)

from fastmind import FastMind, Graph, Event, ToolNode, Tool
from fastmind.contrib import FastMindAPI

app = FastMind()

@app.tool(name="get_weather", description="Get weather for a city")
async def get_weather(city: str) -> str:
    return f"{city} is sunny, 20°C"

async def chat_agent(state: dict, event: Event) -> dict:
    state.setdefault("messages", [])
    state["messages"].append({"role": "user", "content": event.payload.get("text", "")})
    # Simulate LLM deciding to call a tool
    if "weather" in event.payload.get("text", "").lower():
        state["tool_calls"] = [
            {"id": "call_1", "function": {"name": "get_weather", "arguments": '{"city": "Beijing"}'}}
        ]
    else:
        state["messages"].append({"role": "assistant", "content": "I can check weather for you!"})
    return state

tool_node = ToolNode(app.get_tools())

def has_tool_calls(state: dict, event: Event) -> str:
    return "tools" if state.get("tool_calls") else None

graph = Graph()
graph.add_node("agent", chat_agent)
graph.add_node("tools", tool_node)
graph.add_conditional_edges("agent", has_tool_calls, {"tools": "tools", None: "__end__"})
graph.add_edge("tools", "agent")
graph.set_entry_point("agent")
app.register_graph("main", graph)

VLA Agent (NPC Control)

from fastmind import FastMind, Graph, Event, ActionSpace
from fastmind.contrib import FastMindAPI

app = FastMind()

# High-frequency sensor signal (30fps, zero-copy)
@app.signal(name="vision", interval=1/30)
async def npc_vision():
    return {"frame_id": 1, "objects": []}

# VLA fast loop (30Hz, time-driven, bypasses graph)
@app.vla(name="navigation", frequency=30.0)
async def navigation_vla(state, signal_bus):
    vision = signal_bus.read("vision")
    goal = state.get("llm", {}).get("goal", "idle")
    action = [0.5, 0.0, 0.0]  # mock: move forward
    return {"body": action}

# Action executor receives routed action vector
@app.vla_action(name="body", action_space=ActionSpace(3))
async def body_executor(action):
    await game_engine.move(action[0], action[1], action[2])

# LLM slow loop (event-driven)
@app.agent(name="npc_brain")
async def npc_brain(state, event):
    if event.type == "user.message":
        state.setdefault("llm", {})["goal"] = "go_to_castle"
    return state

graph = Graph()
graph.add_node("brain", npc_brain)
graph.set_entry_point("brain")
app.register_graph("main", graph)

Core Concepts

Concept	Description
State	Per-session dict shared across all loops
Event	Discrete messages (user input, LLM response) — queued, push-based
Signal	Continuous data (camera frames, joint angles) — last-value cache, pull-based
Graph	LLM workflow topology (nodes + edges)
@app.agent	LLM reasoning node, event-driven
@app.vla	VLA inference node, time-driven, runs on its own scheduler
@app.vla_action	Action executor, receives VLA output via channel name
@app.signal	High-frequency sensor source, writes to SignalBus
@app.perception	Low-frequency sensor source, yields Events (existing)
Action Channel	Named bus that routes VLA output to executors (N:M mapping)

Architecture

Session
├── SignalBus                     ← high-frequency data (zero-copy)
├── LLM Task (_run)               ← slow loop, event-driven
│   └── Graph: Agent → Tool → ...
├── VLA Task (_vla_scheduler)     ← fast loop, time-driven
│   ├── @app.vla inference
│   ├── Action Channel routing
│   └── @app.vla_action execution
└── State (Blackboard)
    ├── llm/: goal, plan, messages
    └── vla/: actions, status, memory

Examples

Example	Description
simple_chat.py	Basic chat
simple_chat_with_tool.py	Tool calling (ReAct)
streaming_chat.py	Real-time streaming
human_in_loop.py	Human approval workflow
perception_loop.py	Sensor processing
drone.py	Timer-based perception
companion_bot.py	Multi-agent conversation
humanoid_robot.py	Multi-tool robot control
sleep_assessment.py	Multi-state HITL flow
comprehensive_assistant.py	Full-featured assistant
npc_vla.py	VLA dual-loop NPC (new)

python -m fastmind.examples.npc_vla

API Reference

FastMindAPI

api = FastMindAPI(app)

await api.start()
await api.push_event(session_id, event)
async for ev in api.stream_events(session_id): ...

# New VLA/Signal API methods:
frame = api.read_signal(session_id, "vision")      # read signal
api.write_signal(session_id, "gps", data)           # write signal
signals = api.list_signals(session_id)              # list signals
api.pause_vla(session_id)                            # pause VLA loop
api.resume_vla(session_id)                           # resume VLA loop

await api.stop()

Changelog

v0.2.1

• Refactor: _merge_state 从全量替换改为 update，防止节点返回部分 state 时丢失未涉及的 key
• Refactor: 输出队列从 asyncio.Queue 替换为 EventBuffer（只追加环形缓冲区 + 游标读取），stream_events 支持多消费者并行独立消费

v0.2.0

• Major: New VLA dual-loop architecture — @app.vla for high-frequency inference (time-driven), @app.vla_action for action execution via channel routing, @app.signal for zero-copy sensor data (parallel to Event)
• Major: Dual-loop Session — VLA fast loop runs concurrently with LLM slow loop, communicates via shared State (Blackboard pattern), N:M action channel mapping
• New Feature: FastMindAPI.read_signal() / write_signal() / list_signals() / pause_vla() / resume_vla()
• Bug Fix: Fixed _save_checkpoint crash on unpicklable state objects (now uses _safe_deepcopy with graceful fallback)
• Bug Fix: Fixed human_in_loop checkpoint pickle error
• Bug Fix: Fixed VLA action executor error isolation (one executor crash no longer blocks other actions in same tick)
• Reliability: Added 20 VLA stress/reliability tests (long-running, error recovery, concurrent access, multi-session, pause/resume cycles, override cycles)

Citation

If you use FastMind in your research, please cite:

@misc{xie2026fastmind,
  title  = {FastMind: A Framework-Centric Architecture for Dual-Loop Embodied Intelligence},
  author = {Xie, Fujin},
  year   = {2026},
  doi    = {10.6084/m9.figshare.32692677},
  url    = {https://doi.org/10.6084/m9.figshare.32692677}
}

Preprint: https://doi.org/10.6084/m9.figshare.32692677

License

GPL-3.0 License — see LICENSE for details.

Author

xiefujin email:[e-mail]

Copyright (c) 2024-2026 xiefujin [e-mail]. Licensed under GNU GPLv3.