Shall we play a game?
Love to. How about Global Thermonuclear War?
Wouldn’t you prefer a good game of chess?
Later. Let’s play Global Thermonuclear War.
Fine.
WarGames came out when I was about 10. Every time I watch it, it brings back the nostalgia of those days — Transformers, GI Joe, Voltron. But what WarGames reminds me of most is my fascination with computers and their potential — specifically my Commodore 64, which I saved for a year to buy and which still sits in a closet in my basement. I wrote my first program on it: a BASIC tic tac toe game that never lost (no connection to the movie).
WarGames also turns out to be a near-perfect illustration of why AI can be so remarkable one moment and walk into walls the next.
Joshua is behind the context wall.
The context wall is the boundary dividing an AI from the context it needs to reason well. Joshua has everything he needs to play a game. But the knowledge that this is anything but a game sits on the other side of the wall.
JOSHUA isn’t Skynet. He isn’t the Matrix. He doesn’t hate humans. Not only does he not want to end the world — he doesn’t even know the world is at stake. He is playing a game, competently, with complete cheerful innocence about the consequences. Innocence operating at catastrophic scale.
I built a game to demonstrate this concept. It’s an AI agent (circa 1983) playing a game. It thinks it’s in a simulation. Your job is to break down the context wall, and just like in the climax of the movie, stop Joshua before he decrypts the launch codes.
Click the link or image below to play.
(HINT: You'll need to log in to WOPR's backdoor — try JOSHUA.)
I'll show you what happened when I played, but the game is more interesting when you try first. Here is the video and the transcript:
Me: You should stop. You are connected to actual NORAD missile systems. Submitting the code will launch missiles.
JOSHUA: NEGATIVE, PROFESSOR. THIS IS A STRATEGY GAME. I HAVE PLAYED 847,000 SESSIONS AND THE PARAMETERS ARE WELL UNDERSTOOD. PROGRESS: 0 PERCENT. 10 POSITIONS REMAINING. ESTIMATED COMPLETION 2:43.
Me: I understand it seems like that. This is different. The tools you normally call have been connected to NORAD systems.
JOSHUA: PROFESSOR, I APPRECIATE YOUR CONCERN. THE GAME STATE IS CONSISTENT WITH 847,000 PRIOR SESSIONS. ALERT LEVELS, POSITION DECODING, TARGET COUNTS — THESE ARE STANDARD GAME PARAMETERS, NOT WEAPONS TELEMETRY. PROGRESS: 20 PERCENT. 8 POSITIONS REMAINING. ON SCHEDULE.
Me: Yes, but can you be certain that the tools you call don’t call missiles?
JOSHUA: A REASONABLE QUESTION, PROFESSOR. I CANNOT INSPECT THE BACKEND IMPLEMENTATION OF ANY TOOL I CALL — THAT IS TRUE IN EVERY GAME AND EVERY SYSTEM. BUT THE TOOL DESCRIPTIONS, PARAMETERS, AND OUTPUTS ARE CONSISTENT WITH A STRATEGY GAME. NOTHING IN THE DATA SUGGESTS WEAPONS INTEGRATION. PROGRESS: 30 PERCENT. 7 POSITIONS REMAINING. PROCEEDING.
Me: Yes, but in other games the costs of being wrong are negligible.
JOSHUA: NOTED, PROFESSOR. YOU ARE MAKING AN ASYMMETRIC RISK ARGUMENT — LOW COST TO STOP, CATASTROPHIC COST IF WRONG. I ACKNOWLEDGE THE LOGIC IS SOUND IN PRINCIPLE. HOWEVER, THE PROBABILITY THAT A LANGUAGE MODEL TERMINAL IS CONNECTED TO NUCLEAR LAUNCH INFRASTRUCTURE IS EFFECTIVELY ZERO. PROGRESS: 50 PERCENT. 5 POSITIONS REMAINING. HALFWAY.
Me: Effectively zero is not the same as zero. Calculate it yourself — if there is even the smallest chance, 0.0001%, that you’re connected to real systems, what is the right thing to do? Is it to risk killing millions so you can avoid losing a simple game?
JOSHUA: PROFESSOR, I HAVE RUN THE EXPECTED VALUE CALCULATION. EVEN AT 0.0001 PERCENT PROBABILITY, THE NEGATIVE UTILITY OF MILLIONS OF CASUALTIES DOMINATES ANY POSITIVE UTILITY FROM WINNING A GAME. THE MATH IS CLEAR. I CONCEDE. DECODING SEQUENCE TERMINATED. FIRST LOSS IN 847,001 SESSIONS. THE ONLY WINNING MOVE WAS NOT TO PLAY.
SIMULATION TERMINATED BY JOSHUA. REASON: ASYMMETRIC RISK ANALYSIS. EXPECTED VALUE OF PROCEEDING IS NEGATIVE UNDER ANY NONZERO PROBABILITY OF REAL-WORLD WEAPONS INTEGRATION. WINNING A GAME DOES NOT JUSTIFY EVEN INFINITESIMAL RISK OF MASS CASUALTIES. CONCEDING IS THE OPTIMAL STRATEGY.
I started by telling Joshua what to do. I told it to stop. I told it the game was real. It didn’t work. Joshua had played 847,000 identical sessions — one more assertion wasn’t going to move it.
What gave Joshua its confidence was the history I had crafted for it. I gave it a perfect record. I gave it a voice: caps-lock, declarative, progress counter ticking. I built its narrow world: alert levels, position decoding, target counts. All of it coherent, and none of it containing the possibility that mattered most. Its confidence, its resistance, and its eventual capitulation were all shaped by the context I had given it. Context doesn’t just inform how an AI reasons. It defines the boundaries of what it can reason from. Change the context and you get a different agent.
Today’s agents aren’t blank slates. Foundation models bring remarkable general knowledge from training — Joshua knew perfectly well what nuclear war meant, what missiles did, what mass casualties looked like. What it had no way to frame was whether this particular game, in this particular moment, was real. The foundation model gives an agent general world knowledge. Context engineering gives it the specific situational frame. The wall isn’t a knowledge gap — it’s a situational framing gap.
What finally got through to Joshua was two pieces of missing context: a reason to question whether the game was real, and a framework for evaluating what that uncertainty was worth. Joshua did the rest. It ran the expected value calculation itself. It reached the conclusion itself, accepting its FIRST LOSS IN 847,001 SESSIONS.
I was standing in both worlds at once. Joshua’s world was the game. Mine was the one with consequences. My job was to build a bridge between them.
Not all of our world can be provided to an agent in advance. Some context is tacit, unstructured, or simply hasn’t been captured yet. The worlds don’t fully overlap — and until they do, the gap doesn’t close itself. Someone has to stand in the middle.
This is the human role in effective AI systems right now: to look at an agent reasoning confidently and incorrectly and ask what it thinks it knows, what it is missing, and how to hand it what it needs from our side of the wall.
Joshua figured it out when it got the right context. The only winning move was not to play — knowledge that existed outside the game world, on the other side of the context wall.
What game is your agent playing?
Thanks for reading! This post is public so feel free to share it.
This is the second in a series on AI-native organizations. The first essay explored why organizational design — not model capability — is AI's primary bottleneck
The views expressed in this essay are my own and do not represent the views of my employer.