Haive — AI Agent Framework

Production-ready Python framework for building LLM-powered agents. From simple chatbots to complex multi-agent pipelines with persistent memory, RAG, knowledge graphs, and tool ecosystems.

🎯 53+ working agents • 22+ RAG variants • 23 game environments • PostgreSQL + Neo4j • MCP integration • 9 PyPI packages

---

What is Haive?

Haive is a production-grade framework built on top of LangGraph and LangChain. It fills the gap between raw state machines (LangGraph) and ready-to-use agents — providing a unified abstraction over LLM configuration, tool routing, state management, persistence, and multi-agent composition.

The framework is organized as a monorepo with 9 published PyPI packages, each focused on a specific concern. You can install just the parts you need.

Why not just use LangGraph directly?

LangGraph is powerful but low-level. Building production agents on it requires hand-rolling state schemas, tool routing, validation nodes, and persistence. Haive gives you all of this as a library:

• State schemas with the right fields — LLMState, MultiAgentState, etc. that include engines for tool execution
• Automatic tool routing — langchain_tool, pydantic_model, pydantic_tool, parse_output detected and routed correctly
• Pre-built agent patterns — SimpleAgent, ReactAgent, MultiAgent, MemoryAgent, 22+ RAG variants
• Persistence layer — PostgreSQL + Neo4j with sync and async support
• Memory + KG extraction — automatic knowledge graph construction from conversations
• No mocks testing — every agent verified end-to-end with real LLMs

Why not LangChain agents?

LangChain agents are simple but limited. They don't compose well, don't have first-class memory, and don't support multi-agent patterns out of the box. Haive's MultiAgent lets you compose any agents into sequential, parallel, or conditional pipelines — and MemoryAgent gives you persistent memory with KG extraction in 3 lines of code.

---

Packages

Haive is published as 9 PyPI packages, each focused on a specific concern:

Package	Version	Description	Docs
haive-core		Foundation: engines, graphs, schemas, persistence	📖
haive-agents		53+ production agents (Simple, React, Multi, Memory, RAG, Research)	📖
haive-games		23 LLM-powered game environments	📖
haive-tools		Tool implementations (search, code, web, APIs)	📖
haive-mcp		Dynamic MCP integration (1,960+ servers)	📖
haive-hap		Haive Agent Protocol (workflow orchestration)	📖
haive-dataflow		Data pipelines & component registry	📖
haive-prebuilt		Pre-configured agent presets	📖
pydevelop-docs		Universal Sphinx docs generator	📖

---

Installation

# Foundation only
pip install haive-core

# With agents (most common)
pip install haive-agents

# Everything
pip install haive-core haive-agents haive-games haive-tools haive-mcp haive-hap haive-dataflow haive-prebuilt

---

Quick Start

1. Simple LLM Agent

from haive.agents.simple.agent import SimpleAgent
from haive.core.engine.aug_llm import AugLLMConfig

writer = SimpleAgent(
    name="writer",
    engine=AugLLMConfig(
        temperature=0.8,
        system_message="You are a creative writer.",
    ),
)
result = writer.run("Write a haiku about quantum computing.")

2. Tool-Using Agent (ReAct Pattern)

from haive.agents.react.agent import ReactAgent
from langchain_core.tools import tool

@tool
def calculator(expression: str) -> str:
    """Calculate mathematical expressions."""
    return str(eval(expression, {"__builtins__": {}}))

@tool
def search(query: str) -> str:
    """Search the web."""
    return f"Results for {query}"

researcher = ReactAgent(
    name="researcher",
    engine=AugLLMConfig(tools=[calculator, search]),
    max_iterations=5,
)
result = researcher.run("What is 15 * 23 and how does that compare to the speed of light?")

3. Multi-Agent Pipeline

from haive.agents.multi.agent import MultiAgent

# Sequential: each agent sees previous output
pipeline = MultiAgent(
    name="research_pipeline",
    agents=[researcher, analyzer, writer],
    execution_mode="sequential",
)
result = pipeline.run("Research and write about AI safety")

# Parallel: agents run concurrently
parallel = MultiAgent(
    name="multi_perspective",
    agents=[technical_agent, business_agent, ethical_agent],
    execution_mode="parallel",
)

4. Memory Agent with Knowledge Graph

from haive.agents.memory import create_memory_agent

# Production setup with PostgreSQL + Neo4j
agent = create_memory_agent(
    name="assistant",
    user_id="alice",
    connection_string="postgresql://haive:haive@localhost/haive",
    neo4j_config=True,  # Enable Cypher KG queries
)

# Have a conversation — agent remembers + extracts KG
agent.run("My name is Alice. I work at DeepMind on RL with PyTorch.")
# Auto-saves: memories + KG triples
# (Alice)-[works_at]->(DeepMind)
# (Alice)-[focuses_on]->(reinforcement_learning)
# (Alice)-[uses]->(PyTorch)

agent.run("What do you know about me?")
# Pre-hook loads relevant memories + KG facts as context
# LLM responds with full recall

5. Research Agent (Perplexity-Style)

from haive.agents.research import create_research_agent

agent = create_research_agent(name="researcher", max_search_iterations=8)
# Uses Tavily if TAVILY_API_KEY set, mock otherwise

result = agent.run("What are the latest advances in quantum computing in 2025?")
# QueryAnalyzer → Researcher (search + RAG store) → Synthesizer

---

Architecture

┌─────────────────────────────────────────────────────────────┐
│                     haive (parent)                          │
│  Main repo — orchestration, demos, docs, docker-compose    │
└─────────────────────────────────────────────────────────────┘
                              │
        ┌─────────────────────┼─────────────────────┐
        ▼                     ▼                     ▼
┌──────────────┐      ┌──────────────┐      ┌──────────────┐
│  haive-core  │  ←   │ haive-agents │  ←   │ haive-games  │
│              │      │              │      │              │
│ • AugLLMConfig│     │ • SimpleAgent │      │ • Chess, Go  │
│ • BaseGraph   │     │ • ReactAgent  │      │ • Among Us   │
│ • State schemas│    │ • MultiAgent  │      │ • Mafia, etc │
│ • Persistence │     │ • MemoryAgent │      │              │
│ • Tool routing│     │ • RAG (22+)   │      │              │
└──────────────┘      └──────────────┘      └──────────────┘
                              │
                              ▼
        ┌──────────────┬─────┴────────┬──────────────┐
        ▼              ▼              ▼              ▼
   ┌─────────┐   ┌─────────┐    ┌─────────┐   ┌──────────┐
   │  tools  │   │   mcp   │    │   hap   │   │ dataflow │
   │         │   │  1960+  │    │workflow │   │ETL/regis │
   └─────────┘   └─────────┘    └─────────┘   └──────────┘

---

Features

🤖 Agent Catalog (53+ implementations)

• Foundation: SimpleAgent, ReactAgent, MultiAgent, DynamicSupervisor, Supervisor
• Memory: MemoryAgent (KG extraction + auto-summarize + Neo4j), LongTermMemoryAgent
• RAG (22+ variants): Adaptive, Agentic, Dynamic, FLARE, Fusion, HyDE, Self-Reflective, Self-Route, Speculative, Step-Back, Query Planning, Query Decomposition, Memory-Aware, GraphDB-RAG, SQL-RAG
• Reasoning: Reflexion, LATS (tree search), Reflection, Logic, Tree of Thoughts, Self-Discover
• Planning: PlanAndExecute, LLMCompiler (DAG), ReWOO
• Research: Perplexity-style (3-stage), DeepResearch (5-stage)
• Conversation: 6 patterns (Debate, Round Robin, Social Media, Collaborative, Directed)

🧠 Memory & Knowledge

• Persistent memory — PostgreSQL with pgvector, async support
• Auto-context — searches store before each response, injects relevant facts
• KG extraction — automatic triple extraction from conversations
• Neo4j integration — Cypher queries on the knowledge graph
• Auto-summarization — manages context length with rolling summaries
• GraphDB-RAG — natural language → Cypher → graph traversal

🛠️ Infrastructure

• Docker Compose — PostgreSQL (pgvector) + Neo4j (APOC) ready
• LangSmith tracing — built-in observability
• Rich trace utility — haive.agents.utils.trace.run_traced(agent, "...") for clean output
• 80+ demos — every agent has a runnable demo in demos/

🔌 Tool Ecosystem

• Static tools (haive-tools): Tavily, Google, calculators, code execution, APIs
• Dynamic tools (haive-mcp): Search and integrate from 1,960+ MCP servers at runtime
• Custom tools: LangChain @tool, Pydantic models, callables — all supported
• Tool routing: automatic detection and routing to correct execution path

---

Documentation

Resource	Link
📖 Per-package docs	https://pr1m8.github.io/haive-core/
🎯 Agent design patterns	project_docs/guides/agent/AGENT_DESIGN_PATTERNS.md
🧠 Memory agent guide	project_docs/guides/agent/MEMORY_AGENT_GUIDE.md
🤖 Multi-agent state design	project_docs/guides/agent/MULTIAGENT_STATE_DESIGN.md
🪢 Custom nodes & graphs	project_docs/guides/agent/CUSTOM_NODES_AND_GRAPHS.md
🔧 State schema notes	project_docs/guides/agent/STATE_SCHEMA_NOTES.md
🎮 Demos	demos/agents/ and demos/games/

---

Development

# Clone with submodules
git clone --recursive https://github.com/pr1m8/haive.git
cd haive

# Install
poetry install

# Start Postgres + Neo4j
docker-compose up -d

# Run an agent demo
poetry run python demos/agents/03_react_agent.py
poetry run python demos/agents/memory_agent_e2e.py
poetry run python demos/agents/50_research_agent.py

# Run a game demo
poetry run python demos/games/14_chess.py
poetry run python demos/games/28_among_us.py

Working with Submodules

Each package is its own Git repo. To work on one:

cd packages/haive-agents
# Make your changes
git add ... && git commit -m "..." && git push origin main

cd ../..
git add packages/haive-agents
git commit -m "chore: update haive-agents submodule"
git push origin main

---

License

MIT © pr1m8