haive-agents

Production-ready agent implementations for the Haive framework.

53+ working agent implementations covering conversation, planning, reasoning, RAG, memory, research, and multi-agent coordination — built on haive-core and ready for production. Every agent is verified end-to-end with real LLM calls (no mocks).

---

Why haive-agents?

Building production agents from scratch on LangGraph is hard. You need to:

• Roll your own state schemas with the right fields for tool execution
• Implement reasoning loops with iteration tracking and convergence checks
• Handle structured output validation, tool routing, and error recovery
• Wire up memory, persistence, and KG extraction
• Compose agents into pipelines without losing type safety

haive-agents gives you all of this as a library. Each agent is a pydantic.BaseModel you can configure, compose, and extend. Agents follow a consistent pattern: configure with AugLLMConfig, run with agent.run(input), get back structured output.

---

Foundation Agents

SimpleAgent — Conversation + Structured Output

The base agent. Single LLM call, optional structured output, no tools. Use this for chatbots, formatters, classifiers, and any task that doesn't need tool execution.

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.")
print(result.messages[-1].content)

Graph: START → agent_node → END

Use cases: Conversational chatbots, structured data extraction, content generation, classification, summarization.

ReactAgent — Reasoning Loops with Tools

Implements the ReAct pattern: reason → act → observe → repeat. The agent decides which tools to call, sees the results, and continues reasoning until it has an answer or hits the iteration limit.

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 web_search(query: str) -> str:
    """Search the web."""
    return f"Results for: {query}"

researcher = ReactAgent(
    name="researcher",
    engine=AugLLMConfig(
        tools=[calculator, web_search],
        system_message="Use tools to answer questions. Reason step by step.",
    ),
    max_iterations=5,
)

result = researcher.run("What is the population density of Tokyo?")

Graph: START → agent_node → [tool_calls?] → tool_node → agent_node → ... → END

Use cases: Research, math, web scraping, multi-step problem solving, anything that needs tools.

MultiAgent — Compose Agents

Compose multiple agents into pipelines: sequential, parallel, or conditional. Each child agent can be a different type (Simple, React, Memory, etc.). Engines are passed through automatically so child tools work.

from haive.agents.multi.agent import MultiAgent

# Sequential: each agent sees output of previous
pipeline = MultiAgent(
    name="research_pipeline",
    agents=[researcher, analyzer, writer],
    execution_mode="sequential",
)

# Parallel: all run concurrently, results merged
parallel = MultiAgent(
    name="multi_perspective",
    agents=[technical_analyst, business_analyst, security_analyst],
    execution_mode="parallel",
)

# Conditional: route based on state
router = MultiAgent(
    name="router",
    agents=[classifier, simple_handler, complex_handler],
    execution_mode="conditional",
)

Use cases: Research pipelines, multi-perspective analysis, content workflows, fan-out/fan-in patterns.

DynamicSupervisor — Runtime Agent Management

A ReactAgent with handoff tools that can dynamically delegate to other agents. Add agents at runtime, remove them, or create new ones from system messages.

from haive.agents.dynamic_supervisor.agent import DynamicSupervisor

supervisor = DynamicSupervisor(
    name="team_lead",
    engine=AugLLMConfig(system_message="You coordinate a team."),
)

# Add agents at any time
supervisor.add_agent(math_agent, description="Handles math problems")
supervisor.add_agent(writer_agent, description="Writes content")

# Or create one on the fly
supervisor.create_agent(
    name="translator",
    system_message="You translate between languages.",
    description="Handles translation tasks",
)

# Supervisor decides who handles each request
result = supervisor.run("Translate 'hello world' to French")

Use cases: Customer service routing, dynamic task delegation, agent marketplaces, runtime adaptation.

---

MemoryAgent — Persistent Memory + Knowledge Graphs

The most sophisticated agent in the framework. A ReactAgent extended with:

1. Persistent memory — saves facts about users in a store (PostgreSQL or in-memory)
2. Auto-context loading — searches memory before each response, injects relevant facts
3. KG extraction — automatically extracts subject-predicate-object triples from conversations
4. Auto-summarization — summarizes long conversations to manage context length
5. Neo4j integration — optional graph database for Cypher queries on the KG

from haive.agents.memory import create_memory_agent

# Dev mode (in-memory)
agent = create_memory_agent(name="assistant", user_id="alice")

# Production (PostgreSQL with pgvector)
agent = create_memory_agent(
    name="assistant",
    user_id="alice",
    connection_string="postgresql://haive:haive@localhost/haive",
)

# With Neo4j knowledge graph
agent = create_memory_agent(
    name="assistant",
    user_id="alice",
    connection_string="postgresql://haive:haive@localhost/haive",
    neo4j_config=True,  # Uses NEO4J_URI/USER/PASSWORD env vars
)

# Have a conversation — agent remembers and extracts KG
agent.run("My name is Alice. I work at DeepMind on reinforcement learning.")
# → Saves memory: "Alice works at DeepMind"
# → Extracts triples: (Alice)-[works_at]->(DeepMind), (Alice)-[focuses_on]->(RL)
# → Stores in PostgreSQL + syncs to Neo4j

agent.run("I also use PyTorch and JAX.")
# → Saves memory + new triples

agent.run("What do you know about me?")
# → Pre-hook: searches memories + KG triples + summaries
# → Injects context into system message
# → LLM responds with full recall

Memory tools available to the LLM:

• save_memory(content, importance) — save a fact
• search_memory(query) — search past memories
• save_knowledge(subject, predicate, object_) — save a KG triple
• search_knowledge(query) — search KG triples
• query_knowledge_graph(question) — Cypher query (when Neo4j connected)

Document-level KG extraction:

# Extract KG from a document using GraphTransformer
triples = agent.extract_kg_from_document(
    "Marie Curie was born in Warsaw, Poland in 1867. She won two Nobel prizes.",
    allowed_nodes=["Person", "Location", "Award"],
)
# → [{subject: "Marie Curie", predicate: "born_in", object: "Warsaw"}, ...]

Use cases: Personal assistants, customer support with history, research assistants, anything that needs long-term memory.

---

RAG Agents — 22+ Variants

Every meaningful RAG variant from the literature, implemented and ready to use:

Agent	Pattern	Best For
BaseRAGAgent	Simple retrieve → generate	Baseline RAG
AdaptiveRAGAgent	Routes by query complexity	Mixed query types
AgenticRAGAgent	ReactAgent with retrieval tools	Multi-step research
DynamicRAGAgent	Multi-source dynamic retrieval	Varied data sources
FLARERAGAgent	Forward-Looking Active REtrieval	Long-form generation
RAGFusionAgent	Reciprocal rank fusion	Better recall
HyDERAGAgent	Hypothetical document embeddings	Better semantic search
SelfReflectiveRAGAgent	Generate → grade → regenerate	Hallucination control
SelfRouteRAGAgent	Query-aware route selection	Mixed strategies
SpeculativeRAGAgent	Hypothesis + parallel verification	Speed + accuracy
StepBackRAGAgent	Abstract query → retrieve → answer	Complex reasoning
QueryPlanningRAGAgent	Query planning + multi-hop	Hierarchical questions
QueryDecomposerAgent	Decompose into sub-queries	Compound questions
MemoryAwareRAGAgent	RAG + memory context	Personalized RAG
GraphDBRAGAgent	NL → Cypher → Neo4j	Graph-based knowledge
SQLRAGAgent	NL → SQL → database	Structured data

from haive.agents.rag.adaptive.agent import AdaptiveRAGAgent
from langchain_core.documents import Document

docs = [Document(page_content="...") for _ in range(100)]

agent = AdaptiveRAGAgent.from_documents(
    documents=docs,
    llm_config=AugLLMConfig(),
    max_query_complexity="complex",
)

result = agent.run("What are the key insights from the documents?")

---

Reasoning & Critique Agents

Implementations of recent reasoning paper algorithms:

Reflexion — Draft → Reflect → Revise

from haive.agents.reasoning_and_critique.reflexion.agent import ReflexionAgent

agent = ReflexionAgent(
    name="reflexive",
    engine=AugLLMConfig(),
    max_revisions=3,
)
result = agent.run("Solve this complex problem...")
# → Draft → Reflect on draft → Revise → Loop until quality threshold

LATS — Language Agent Tree Search

Tree search with UCB1 selection, simulation, backprop, and reflection-based scoring.

from haive.agents.reasoning_and_critique.lats.agent import LATSAgent

agent = LATSAgent(
    name="lats",
    engine=AugLLMConfig(tools=[search]),
    max_depth=4,
    n_samples=3,
)

Other reasoning agents: ReflectionAgent, LogicReasoningAgent, ToTAgent (Tree of Thoughts), SelfDiscoverAgent.

---

Planning Agents

PlanAndExecuteAgent — Planner → Executor → Replanner

from haive.agents.planning.plan_and_execute import PlanAndExecuteAgent

agent = PlanAndExecuteAgent.create(
    tools=[calculator, web_search],
    name="planner",
)
result = agent.run("Research AI safety and write a 3-paragraph summary.")

LLMCompilerAgent — DAG-based Parallel Execution

Plans tasks as a DAG, executes independent tasks in parallel, joins results, replans if needed.

from haive.agents.planning.llm_compiler.agent import LLMCompilerAgent

agent = LLMCompilerAgent(
    name="dag_executor",
    engine=AugLLMConfig(tools=[search, calc]),
)

ReWOO — Reasoning Without Observation

Plans all steps upfront, executes them all, then synthesizes. Only 2 LLM calls for the planning + synthesis (intermediate steps don't require LLM reasoning).

from haive.agents.planning.rewoo.agent import ReWOOAgent

agent = ReWOOAgent(name="rewoo", engine=AugLLMConfig(tools=[...]))

---

Research Agents

ResearchAgent — Perplexity-Style 3-Stage

QueryAnalyzer → Researcher (search + RAG) → Synthesizer

from haive.agents.research import create_research_agent

agent = create_research_agent(
    name="research",
    max_search_iterations=8,
    # Tavily auto-detected if TAVILY_API_KEY set, else mock
)
result = agent.run("What are the latest advances in quantum computing?")

DeepResearchAgent — 5-Stage Pipeline

Planner → Researcher → Analyzer → FactChecker → Writer

Shared ResearchStore across agents so the analyzer can retrieve what the researcher found.

from haive.agents.research import create_deep_research_agent

agent = create_deep_research_agent(
    name="deep",
    include_fact_check=True,
    max_search_iterations=10,
)
result = agent.run("Compare React vs Vue.js for 2025 web development")

---

Conversation Agents

Six conversation patterns:

Agent	Description
BaseConversationAgent	Foundation for all conversation types
CollaborativeConversation	Agents collaborate on a goal
DebateConversation	Structured debate with positions
DirectedConversation	Moderator-directed flow
RoundRobinConversation	Sequential turn-taking
SocialMediaConversation	Social media simulation with personalities

from haive.agents.conversation.debate.agent import DebateConversation

debate = DebateConversation(
    name="ai_safety_debate",
    engine=AugLLMConfig(),
    topic="Should AI development be paused?",
    n_rounds=3,
)
result = debate.run("Begin debate")

---

Trace Utility — Pretty-Print Any Agent

Strip the noisy LangGraph debug output and see clean agent execution:

from haive.agents.utils.trace import run_traced

result = run_traced(agent, "Tell me about quantum computing", save_to="traces/")
# → Rich-formatted tree showing user message, AI response, tool calls,
#   tool results, timing, token usage, store contents

---

Installation

pip install haive-agents

For specific extras:

pip install haive-agents[memory]    # MemoryAgent + Neo4j support
pip install haive-agents[rag]       # RAG dependencies
pip install haive-agents[research]  # Tavily + research deps

---

Quick Start

from haive.agents.simple.agent import SimpleAgent
from haive.agents.react.agent import ReactAgent
from haive.agents.multi.agent import MultiAgent
from haive.agents.memory import create_memory_agent
from haive.core.engine.aug_llm import AugLLMConfig
from langchain_core.tools import tool

# 1. Simple LLM agent
writer = SimpleAgent(
    name="writer",
    engine=AugLLMConfig(temperature=0.8, system_message="You are a writer."),
)

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

researcher = ReactAgent(
    name="researcher",
    engine=AugLLMConfig(tools=[search], system_message="Use search tool."),
    max_iterations=3,
)

# 3. Compose into pipeline
pipeline = MultiAgent(
    name="research_pipeline",
    agents=[researcher, writer],
    execution_mode="sequential",
)
result = pipeline.run("Write an article about quantum computing")

# 4. Add persistent memory
memory_agent = create_memory_agent(
    name="assistant",
    user_id="user123",
    connection_string="postgresql://haive:haive@localhost/haive",
)
memory_agent.run("My name is Alice and I work at DeepMind on RL.")
memory_agent.run("What do you know about me?")  # Recalls everything

---

Documentation

📖 Full documentation: https://pr1m8.github.io/haive-agents/

---

Related Packages

Package	Description
haive-core	Foundation: engines, graphs, schemas
haive-tools	Tool implementations
haive-games	LLM-powered game agents
haive-mcp	Dynamic MCP integration

---

License

MIT © pr1m8