The Fragmented Agent Crisis: Orchestrating Tool Silos with Meta-Harness Architecture

This file, supervisor.py, uses the LangGraph-supervisor pattern. It is a normal Python application.

# supervisor.py — user-defined multi-agent supervisor
from langchain_openai import ChatOpenAI
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_community.agent_toolkits.sql.toolkit import SQLDatabaseToolkit
from langchain_community.utilities.sql_database import SQLDatabase
from langchain_core.vectorstores import InMemoryVectorStore
from langchain_openai import OpenAIEmbeddings
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_community.document_loaders import PyPDFLoader
from langgraph.prebuilt import create_react_agent
from langgraph_supervisor import create_supervisor

# ---- Shared LLM (configurable per team model policy) ----
llm = ChatOpenAI(model="gpt-4o", temperature=0)

# ---- Sub-agent 1: Research (web) ----
research_agent = create_react_agent(
    model=llm,
    tools=[TavilySearchResults(max_results=5)],
    name="research_agent",
    prompt="""You are the research agent.
Use Tavily to search the public web for context, markets, and competitors.
Always cite sources by URL. Never make up statistics without a source.""",
)

# ---- Sub-agent 2: Finance (internal SQL) ----
db = SQLDatabase.from_uri("sqlite:///./finance.db")
sql_toolkit = SQLDatabaseToolkit(db=db, llm=llm)
finance_agent = create_react_agent(
    model=llm,
    tools=sql_toolkit.get_tools(),
    name="finance_agent",
    prompt="You answer revenue, ARR, churn and unit-economics questions. "
                  "Read-only. Never UPDATE/INSERT/DELETE. Return row counts and SQL.",
)

# ---- Sub-agent 3: Compliance (RAG over policy corpus) ----
docs = PyPDFLoader("./policies/company_handbook.pdf").load()
chunks = RecursiveCharacterTextSplitter(chunk_size=800, chunk_overlap=80).split_documents(docs)
vs = InMemoryVectorStore.from_documents(chunks, OpenAIEmbeddings())
retriever = vs.as_retriever(search_kwargs={"k": 4)

from langchain.tools import tool

@tool
def policy_search(query: str) -> str:
    """Search the internal policy corpus for compliance clauses."""
    hits = retriever.invoke(query)
    return "\n\n".join(d.page_content for d in hits)

compliance_agent = create_react_agent(
    model=llm,
    tools=[policy_search],
    name="compliance_agent",
    prompt="You interpret internal policy. Cite the chunk you used (filename + page). "
                  "If a question is outside policy scope, say so and recommend escalation.",
)

# ---- The supervisor graph ----
supervisor = create_supervisor(
    agents=[research_agent, finance_agent, compliance_agent],
    model=llm,
    prompt="""You are the enterprise analysis supervisor.
Route the request to the right specialist:
  - research_agent    → public web, market, competitive intel
  - finance_agent     → revenue, ARR, churn, unit economics (read-only SQL)
  - compliance_agent  → policy, legal, regulatory interpretation
If a request crosses domains, decompose it into ordered sub-routes.
Always return a structured summary with citations per sub-agent.""",
).compile(
    checkpointer=True  # durable execution via LangGraph checkpointing
)

This is production-shaped LangGraph code. It compiles to a runnable graph with checkpointed state. It could be deployed to LangSmith Deployment today. However: when the data team wants the marketing team to collaborate on a session, there is no shared workspace. When CISO changes the policy corpus, there is no central policy plane. When finance asks, “who spent what”, there is no cross-agent cost aggregation.