partner-onboarding-agent

LangGraph agent that takes a freshly signed B2B partner from contract to operational in under 2 hours. Zero manual handoffs between sales, ops, and support.

A single closed-won HubSpot deal becomes: a provisioned M365 mailbox + SharePoint site + Planner board, a Zendesk org with the right SLA tier, an internal portal account, an AI-generated welcome email sequence + kickoff agenda + 30-day readiness checklist, and a co-branded PDF welcome packet — uploaded to both the portal and SharePoint, with a summary note posted back to the deal.

Co-branded welcome packet (auto-generated per partner)

Verified against a real Microsoft 365 tenant

Live readback from the production tenant, after the agent ran end-to-end with B2B_USE_MOCKS=false. Mailbox, M365 group, SharePoint site, Planner board (with all four buckets) and the uploaded PDF are all real Graph resources, addressable by their IDs in the screenshot below.

What it does

When a HubSpot deal flips to Closed Won, the agent:

1. Reads partner profile (tier, region, services purchased) from HubSpot
2. Provisions a shared mailbox in Microsoft 365
3. Creates a SharePoint site (via M365 group) for shared documents
4. Spins up a Planner board with onboarding buckets
5. Creates a Zendesk organization + attaches the right SLA policy for the tier
6. Provisions an account in the internal portal (mocked here) and an intake form
7. Generates a co-branded PDF welcome packet (HTML + WeasyPrint) and uploads it
8. Generates a 5-touch welcome email sequence + kickoff call agenda + 30/60/90 readiness checklist
9. Posts a summary note back to the HubSpot deal

State is checkpointed via LangGraph's SQLite checkpointer so a stalled run can be resumed.

Architecture

HubSpot webhook ──► FastAPI receiver ──► enqueue ──► LangGraph run
                                                          │
                          ┌───────────────────────────────┴───────────────────┐
                          ▼                                                   ▼
                read_partner_profile                              generate_content (LLM)
                          │                                                   │
                provision_m365 (mailbox → site → planner)             pdf_packet
                          │                                                   │
                provision_zendesk                                      upload_assets
                          │                                                   │
                provision_portal ─────► (FastAPI mock server)
                          │
                post_back_to_hubspot

Quick start

# 1. Install both packages
cd ../b2b-agent-toolkit && pip install -e ".[dev]" && cd -
pip install -e ".[dev]"

cp .env.example .env
# add ANTHROPIC_API_KEY; leave B2B_USE_MOCKS=true for the demo

# 2. Run the mock portal in one terminal
mock-portal

# 3. Trigger an onboarding run in another
onboard run --deal-id deal-001

You'll get a PDF welcome packet at out/<partner>-welcome-packet.pdf and a JSON event log in out/<run-id>.json.

Going to production

• Set B2B_USE_MOCKS=false and fill in the M365 / HubSpot / Zendesk creds in .env
• Replace B2B_PORTAL_BASE_URL with the real portal URL
• Wire the /webhook/hubspot/deal-stage-changed endpoint to your HubSpot account
• Swap the SQLite checkpointer for Postgres in src/onboarding/graph.py

Layout

src/onboarding/
├── state.py            # OnboardingState (LangGraph TypedDict)
├── graph.py            # StateGraph wiring + compile
├── nodes.py            # provisioning nodes (one per integration)
├── content.py          # LLM-generated welcome email, agenda, checklist
├── pdf.py              # Jinja2 + WeasyPrint co-branded packet
├── webhook.py          # FastAPI receiver for HubSpot
└── cli.py              # Typer CLI: `onboard run --deal-id ...`
mock_portal/
├── main.py             # FastAPI mock of the internal portal API
└── store.py            # in-memory state
templates/
└── welcome_packet.html # branded packet template