Microsoft Agent Framework — Multi-Agent (local)

Microsoft Agent Framework is Microsoft’s open-source SDK for building production AI agents in Python and .NET — single agents, multi-agent workflows, and hosted-agent deployment. Neo4j is the graph database and knowledge layer that grounds those agents in connected enterprise data.

This example wires them together as a multi-agent investment-research assistant: a Coordinator delegates to a Database agent (queries Neo4j) and an Analyst (synthesizes a report). One uv-runnable file. Reads microsoft-foundry/.env for the Foundry endpoint and Neo4j credentials.

Tool names and return shapes follow the EXAMPLE_AGENT.md spec ("Industry Research Agent").

When to pick this path

You want to see how Agent Framework composes multiple agents.
A single fat agent with twenty tools loses focus — splitting work into a Database agent and an Analyst produces sharper, more grounded output.
Function tools talk to Neo4j directly via the Bolt driver — no MCP server, no extra hop.

Quick start

az login
cd microsoft-foundry/infra && ./deploy.sh    # one-time, opt in to Foundry
cd ../../microsoft-agent-framework/examples/multi-agent && uv run multi_agent_neo4j.py

uv reads the inline # /// script deps at the top of multi_agent_neo4j.py and runs. The script reads microsoft-foundry/.env for the Foundry endpoint, Azure tenant, and Neo4j credentials.

The three agents

flowchart LR
    user["User"] --> coord
    subgraph agents["Multi-agent system"]
        coord["Coordinator Agent<br/>(delegates)"]
        db["Database Agent<br/>(10 Neo4j functions)"]
        analyst["Analyst Agent<br/>(no tools — synthesis)"]
    end
    coord -->|as_tool| db
    coord -->|as_tool| analyst
    db -->|neo4j Bolt driver| neo4j[("Neo4j Aura<br/>(companies demo graph)")]

Agent Tools Job

Agent	Tools	Job
Coordinator Agent	`database_agent.as_tool()`, `analyst_agent.as_tool()`	Orchestrates: delegates one focused question per facet to the Database Agent, concatenates responses, passes to the Analyst Agent. Never queries the graph or writes the report itself.
Database Agent	10 typed Neo4j functions	Calls the right tool for the request, returns one `json` block per call (raw rows verbatim — every `company_id`, `article_id`, title, date, sentiment, relationship type). No prose.
Analyst Agent	none	Reads the JSON blocks, produces a structured report: Executive Summary, Profile, Recent Developments, Network table, Risks & Outlook. Cites IDs verbatim from the rows.

Coordinator Agent

database_agent.as_tool(), analyst_agent.as_tool()

Orchestrates: delegates one focused question per facet to the Database Agent, concatenates responses, passes to the Analyst Agent. Never queries the graph or writes the report itself.

Database Agent

10 typed Neo4j functions

Calls the right tool for the request, returns one json block per call (raw rows verbatim — every company_id, article_id, title, date, sentiment, relationship type). No prose.

Analyst Agent

none

Reads the JSON blocks, produces a structured report: Executive Summary, Profile, Recent Developments, Network table, Risks & Outlook. Cites IDs verbatim from the rows.

Composition is two Agent.as_tool() calls handed to the Coordinator’s tools= — that’s the whole multi-agent surface. The script is intentionally self-contained; ../foundry-hosted/main.py is a parallel near-identical file packaged for the Foundry hosted-agent runtime.

Function tools (Neo4j)

Ten read-only functions over the public companies demo graph, organised the way the agent picks them — same set as microsoft-foundry/examples/foundry-sdk/:

Discovery — search_companies, list_industries, companies_in_industry Profile — query_company Network — analyze_relationships, people_at_company News — search_news(company_name, query) (vector — embeds query via OpenAIEmbeddingClient against the Foundry text-embedding-3-small deployment, hits the demo graph’s 1536-dim news index), articles_in_month, get_article, companies_in_article

Plain Python functions with type hints and docstrings. Agent Framework auto-converts them to function tools — no decorator, no schema boilerplate.

Anti-hallucination contract

as_tool() only forwards the inner agent’s text to the parent — easy for an inner agent to silently summarise IDs away. Three guarantees in the instructions:

Database Agent emits one fenced json block per tool call (tool, args, rows). No prose, no summaries.
Coordinator passes those JSON blocks verbatim into the Analyst’s task. Never strips down to bare IDs.
Analyst Agent must cite every company_id, article_id, title, and relationship type from the rows. Real IDs look like EIsFKrN_ZNLSWsvxdQfWutQ / ART11195006745 — short placeholders ("101", "AWS partnership") are flagged in the prompt as hallucination.

Result: every value in the report appears verbatim in a tool result.

How it authenticates

You → Foundry: AzureCliCredential pinned to AZURE_TENANT_ID from .env so it works when az login is logged into multiple tenants.
You → Neo4j: the neo4j Python driver with username/password from .env. Defaults to companies / companies against the public demo graph.

No Foundry tokens or Neo4j credentials are passed through the model — it sees only the tool schemas and the rows you return.

Override knobs

Set these in microsoft-foundry/.env:

Variable Default Purpose

Variable	Default	Purpose
`FOUNDRY_QUESTION`	"Research Microsoft’s position…"	The single user question. Pick something that exercises both database and analyst.
`FOUNDRY_MODEL_DEPLOYMENT_NAME`	`gpt-4o-mini`	Model to run all three agents on.
`NEO4J_URI` / `NEO4J_DATABASE` / `NEO4J_USERNAME` / `NEO4J_PASSWORD`	demo graph	Point at your own Aura or self-managed Neo4j.

FOUNDRY_QUESTION

"Research Microsoft’s position…"

The single user question. Pick something that exercises both database and analyst.

FOUNDRY_MODEL_DEPLOYMENT_NAME

gpt-4o-mini

Model to run all three agents on.

NEO4J_URI / NEO4J_DATABASE / NEO4J_USERNAME / NEO4J_PASSWORD

demo graph

Point at your own Aura or self-managed Neo4j.