TaskDrift Onboarding Agent

• 

  Dashboard

• 

  Architecture diagram

What is TaskDrift Onboarding Agent

I run TaskDrift, an AI dev and automation agency. Every new client goes through the same manual grind - scope the project, quote a price, chase them if they go quiet, send a payment link once they agree. I was doing this by hand for every lead, and it does not scale.

So I built an agent system that does the whole thing end to end.

What it does

A multi-agent system on Google ADK. One agent scopes the request, one handles repricing if the client pushes back, one follows up through WhatsApp or email if they go cold, and one generates a Razorpay payment link once things are agreed. No manual steps.

It runs on Gemini for reasoning, Firestore for state between steps, Cloud Run to serve it, and Cloud Scheduler to trigger follow-ups on a timer. Frontend is Next.js 14 with Tailwind.

How I built it

I mapped my own onboarding process step by step and split it into agent responsibilities. Each agent has one job only - scope, pricing, outreach, payment stay separate. This made debugging far easier than one big prompt trying to do everything.

State passes through Firestore so any agent can resume even if a client takes days to reply. Cloud Scheduler handles timing so follow-ups go out without me checking a calendar.

Challenges I faced

Agent handoffs losing context was the hardest part. Sometimes the pricing agent would start before the scope agent's Firestore write finished, so it acted on incomplete data. Had to add checks so an agent will not start until the previous write is confirmed.

WhatsApp Cloud API has strict message windows and formatting rules, so the follow-up agent had to work within those instead of firing on every scheduler tick.

Razorpay needed guarding too - making sure a client replying twice, or a duplicate scheduler fire, does not generate two payment links for the same person.

What I learned

A lot of my own workflow was just repeated decisions I already knew the answer to. Turning that into agent logic instead of one big prompt made the system far more reliable.

What is next

Plug this directly into TaskDrift's live pipeline so every new lead runs through it automatically, not just as a demo.

Built With

• cloud-run
• cloud-scheduler
• firestore
• gemini
• google-adk
• nextjs
• python
• razorpay
• resend
• tailwindcss
• typescript
• vertex-ai
• whatsapp-cloud-api