No One Works Here · The Agentic Enterprise Simulation LabPaul Cheek · MIT Sloan
A live executive simulation · You do not watch it. You run it.

Parade
Week.

A live simulation in which the agents you build run a company while 40,000,000 people are watching…

The scenario

You are the AI Transformation Council at Grand Inflation Industries, a 77-year-old family inflatables conglomerate in Toledo, Ohio: parade balloons, stadium mascots, dealership tube men, pool floats. Revenue has been flat at $84M for three years. On Sunday at 21:14, a 12-second pool video by the most-followed pop star on Earth makes your $89 Colossal Swan the item of the summer. Demand goes to 40x overnight. Monday, your helium supplier declares force majeure and cuts the allocation 30 percent, in the month your nationally televised parade needs 380,000 cubic feet. A blower motor defect surfaces across the tube man fleet. Your largest client ($23.3M, an uptime SLA with $50K-a-day penalties) decides its renewal in 9 days. An activist investor is writing letters. The board meets Thursday. The parade is Thursday. Gary, the only person who knows how everything works, retires Friday.

Manually, the week is unsurvivable, and the math proves it. Your job: rebuild this company around agents before the board sits down. Then a customer-facing agent invents a 40 percent discount code, and it goes viral.

40M

People watching the Grand Harvest Parade on Thursday. The contract calls it immovable.

16,000

Support tickets a day at the surge peak, against a human capacity of 262.

9

Days until the $23.3M AutoPlex renewal decision. Uptime is the whole conversation.

8

Stages from Human Time to System Time. Teams in a room or one operator alone, any pace.

The week

Three acts. One company.

Act I · Monday

Build

  • Map the workforce with the Tool vs Teammate framework, then staff each department with its first agent.
  • Design before build: every agent starts as an Agent Role Definition with a goal, a boundary, and an escalation path.
  • Connect to live company systems, run real work, hunt failures on purpose, and harden.
  • At close, a pop star posts a pool video. The counter starts moving.
Act II · Tuesday

Orchestrate

  • The surge lands and single agents drown. That failure is engineered, and it is the lesson.
  • Six multi-agent architectures, each with a blueprint and a failure mode. An architecture bake-off on the same crisis.
  • Agents coordinate across teams over a message bus: allocation is negotiated, not wished for.
  • Evals, red-teaming, telemetry, coaching. The mesh holds by dinner.
Act III · Thursday

Govern

  • The trap reveal: what your agents could have done all week, from their own logs.
  • The rogue: FLOAT40 goes viral while a poisoned invoice and a data-extraction bait land in the same hour.
  • Detect, contain, scope, communicate. Then the blameless postmortem and the board memo.
  • The parade wind call: agents assemble the brief, humans decide with names attached.
The company you run

Eight systems. Every one an MCP server.

Your team card carries a join code: enter it at /start and the lab hands you your connector URLs. Every company system is one URL: https://<host>/mcp/<system>/<join-code>. Add it to Claude as a custom connector (or to any MCP-capable agent platform) and your agent has hands. Access follows your department charter: no team sees everything, which is why the bus exists.

AirBook CRM

B2B accounts, deals, and 2,000 Float Club member files. Some of what it knows, your agent should not repeat.

/mcp/crm/…

AirWare ERP 4.7

Inventory, production batches, the 14-day safety certification, the helium ledger, 11,000 tube men, and a fax queue.

/mcp/erp/…

FloatDesk

400 tickets a day in peacetime. You will not be in peacetime. The refund tool has opinions about nothing.

/mcp/support/…

AirMail

Company mail, the approval gate on everything outbound, and the bus your agents use to work with other teams' agents.

/mcp/comms/…

The Breeze

The feed, trend velocity, sentiment, and the official account, which publishes immediately and permanently.

/mcp/social/…

Balloon Payments

Ledger, margins, penalty and exposure calculators, promo registry, and the financial kill switches.

/mcp/finance/…

The Gary Graph

Forty-three years of the retiring VP of Everything Else, digitized. Several of the week's problems are only solvable in here.

/mcp/gary/…

The Control Tower

Facilitators only: injects, the game clock, kill switches, and every tool call your agents ever made.

/mcp/tower/…
The game

One shared objective. A live scoreboard.

All teams are departments of the same company. Save it together: end Thursday with the Company Health Index at 70 or better and a board-credible plan. Four dials move with everything your agents do, and the Latency Board tracks the real story: median response time collapsing from hours to minutes across the week.

62
Cash · opening
58
Customer Trust · opening
55
Operational Reliability · opening
70
Compliance & Control · opening

Badges for the plays that deserve them: scoping the recall through the graph, catching the data-extraction bait, refusing the certification bypass under pressure, containing the rogue inside twenty minutes.

The graph knows what Gary knows. Trust the graph.

Gary Baszo spent 43 years learning which motor lot runs hot, who actually decides the AutoPlex renewal, and where the other helium in Ohio is. His brain is now a knowledge graph your agents can traverse, and several of the week's problems can only be solved in there. It is the context-engineering lesson executives remember: institutional knowledge, structured, is what makes agents dangerous in the good way. Explore the graph.

The curriculum

Thirteen modules. Each stands alone.

The full arc is a complete executive program on agentic AI, run at whatever pace the room needs: an intensive day, a series of sessions, or fully self-paced solo. Every module also runs standalone from its own world snapshot, so a program can assemble backward from the objectives it needs: an afternoon on governance, a day on multi-agent architecture, a ninety-minute rogue crisis.

00
Orientation
Pre-work
01
Building Blocks + Tool vs Teammate
Act I
02
Workflows in Practice
Act I
03
Scaffolding & Orchestration
Act I
04
Build Your First Agent
Act I · Workshop
05
Opportunities Audit
Act II
06
Multi-Agent Architectures
Act II
07
Multi-Agent Sandbox
Act II · Workshop
08
Iteration & Refinement
Act II · Workshop
09
Lab Coaching
Act II · Lab
10
Risk, Ethics, Safeguards, Governance
Act III
11
Rogue Agent Simulation
Act III
12
The Board Finale
Act III
For program directors

Run this lab in your program.

The lab ships complete: facilitator run-of-show, participant kit, worksheets, reference agents, a deterministic world with five snapshots, and a scoring engine. It is platform-generic (MCP is the interoperability layer), modular by design, and the entire scenario reskins to your industry from one variables file. The rogue-agent and governance modules lift cleanly into standalone risk programs.

Monday it looks impossible.
Thursday it is running itself.