← back to work

Autonomous Procurement Multi-Agent System

Two AI agent systems that autonomously negotiate and execute end-to-end procurement workflows.

2026 · YearCreator2 monthsOpen Source
ProcuForge
procuforge.png
overview

What it is

ProcuForge is an end-to-end autonomous procurement platform where two independently deployed multi-agent systems negotiate, decide, and execute purchase workflows over the A2A protocol — built on Google ADK and Vertex AI Agent Engine.

challenge

The hard part

Procurement workflows are slow, manual, and full of back-and-forth between teams. Could two AI agent systems handle the entire negotiation and execution loop autonomously — while keeping humans in the loop for critical approval gates?

approach

How I thought about it

I built two independently deployed multi-agent systems (buyer and supplier) that communicate over the A2A protocol. The full stack includes a FastAPI + WebSocket backend, a Next.js UI with real-time workflow updates, a Firestore audit trail, and Cloud Run / Vertex AI Agent Engine deployment — with human-in-the-loop approval gates for PO, GRN, and completion.

stack

Tech

Python
FastAPI
Next.js
Google ADK
Vertex AI
Firestore
Docker
features

What it does well

Multi-agent negotiation

Buyer and supplier agents communicate over A2A protocol, negotiating terms autonomously.

Human-in-the-loop gates

Approval checkpoints for PO, GRN, and completion — humans stay in control of critical decisions.

Real-time UI

Next.js frontend with WebSocket updates showing live agent decisions and workflow state.

Full audit trail

Every agent action and decision logged to Firestore for traceability and compliance.

lessons

What I took away

  • A2A protocol makes multi-agent orchestration composable — agents become services.
  • Human-in-the-loop gates are not a limitation; they are a feature that builds trust.
  • Real-time WebSocket updates transform agentic UX from opaque to transparent.