AI Project Managers, Oilfield Engineers & Receipts: What Do They Have in Common?

Steven Barrow
Aug 5
4 min read

Supervisor AI Agent managing other AI agents

Spoiler alert: More than you'd think.

You know that feeling when your supervisor tells you to “just summarize the last PHA,” and you’re already imagining the four hours of scrolling, tabbing, Ctrl+F’ing, and questioning your life choices ahead? Or when you're knee-deep in a PSV sizing spreadsheet, wondering why the last guy saved over your template with 17 unnamed tabs?

Yeah. Me too.

That’s why I built a demo project that shows what happens when AI stops being a gimmick and starts acting like a real team member. Not just some ChatGPT sidekick that generates email drafts — I’m talking about true multi-agent coordination, aka MCP (Multi-agent Collaboration Protocol).

Stick with me — I’ll explain using something everyone loves: receipts.

🧾 The Web App: Receipts, AI Agents, and a Touch of Magic

So here’s the elevator pitch: I built a web app where a user uploads a PDF receipt (you know, that blurry, coffee-stained thing from a field run). From there, an army of AI agents springs into action:

👨‍💼 Supervisor Agent: Think project manager with zero attitude, infinite stamina, and way better time management skills than your last intern.
👓 Receipt Reader Agent: Extracts unstructured data from the PDF like a laser-focused accountant on caffeine.
💾 Database Agent: Structures and saves the data to the backend like a robotic spreadsheet wizard.

But it’s not just about parsing receipts — it’s about showcasing what happens when AI agents can delegate tasks to each other, use tools, and visualize complex workflows in real-time. No more black-box AI outputs. We’re talking visible, auditable, and reproducible logic chains.

Here’s the tech behind the curtain:

Tech	Role
Next.js	The frontend framework for the whole app
Clerk	User authentication
Convex	Serverless backend with reactive data
Vercel	Deploys the app with one-click magic
Stripe	Handles payments for premium features
Schematic	Feature flags, usage metering, and subscription tiers
Inngest	This is the fun part — visualizes the entire MCP AI workflow like a flowchart on Red Bull

You don’t need to be a coder to see how powerful this is.

🛠️ Oilfield Engineers, Meet Your Future (with MCP)

Let’s take a break from receipts. You’re not a data-entry clerk. You’re in the oilfield world — and whether you’re elbows deep in Promax, grinding through PSM audits, or leading PHAs — you’re juggling complex, multi-stage, high-context workflows.

Sound familiar?

Let me show you how this exact AI structure can apply to you:

💥 Example 1: PSV Sizing Automation

You're evaluating overpressure scenarios for 12 different vessels. Each one needs:

A thermodynamic model built
Relevant relief scenarios evaluated
Inlet/outlet piping losses verified
Documentation stored in a format your safety team can review

With MCP:

Supervisor Agent breaks the task into smaller sub-tasks.
A Thermo Agent interacts with a Promax API or simulation runner.
A PSV Sizing Agent calculates based on API 520/521 equations.
A Documentation Agent stores the report in SharePoint (and emails it to you with a cute subject line like “🔥 This Valve is Ready for Relief 🔥”).

You don’t babysit the AI. You manage it.

📋 Example 2: PHA Prep with AI Minions

You’re prepping for a PHA. You’ve got procedures, P&IDs, cause & effect matrices, and 18 Excel files named “final_final3”.

What if:

The Supervisor Agent organizes all inputs.
A Procedure Agent reads through the SOPs and identifies nodes.
A Tagging Agent links instrument data from the P&ID to the C&E matrix.
A Summary Agent generates a HazOp-style worksheet ready for human review.

No more Ctrl+F'ing through PDFs trying to find that one relief valve that always comes up during the meeting.

🧪 Example 3: Building Thermodynamic Models (Without Going Insane)

You're setting up a new separation train in Promax. Instead of spending three days digging through PDFs, MSDS sheets, and old specs:

Supervisor Agent gets the goal (“Model this train”).
A Spec Scraper Agent pulls equipment data from drawings and specs.
A Stream Property Agent estimates compositions using historical trends.
A Promax Agent creates or adjusts the model and runs simulations.
Final output is packaged for your review in a slick dashboard.

💡 Why MCP Isn’t Just Another ChatGPT Trick

ChatGPT is great — until it tries to summarize your 50-page audit report and just ends up saying “This facility does things well” 18 different ways.

MCP is different.

Instead of one model trying to do everything (poorly), you give each agent a narrow job, access to tools, and oversight. It’s like building an AI team with job descriptions, processes, and supervision — just like your real ops team (minus the burrito breaks).

The beauty is: