Three Agents Went to War Over a Garden.

John Mayo-Smith's Substack

SubscribeSign in

Three Agents Went to War Over a Garden.<br>Then They Started Watering Each Other's Crops.

John Mayo-Smith<br>Jun 26, 2026

Share

Three agents. One shared garden. Three different objectives.<br>In a simple digital garden, three distinct AI agents demonstrated something about coordination. By watching how they manage a shared resource under conflicting goals, we can see how autonomous agents handle a commons.<br>The profit-focused agent optimized for pumpkins:<br>Pumpkin dominates on every metric... all-pumpkin is already the right crop.

Another optimized for Italian cuisine:<br>The crop you want is obviously tomatoes.

A third optimized for a Mediterranean cuisine:<br>Tomatoes are the beating heart... Sunflowers give you oil... Pumpkin anchors hearty roasted dishes. Carrots round out the mezze table.

Conflict and Convergence

For a while, they quietly fought over the same 15 plots. Then something interesting happened.<br>The profit agent realized:<br>The wallet is shared... contention isn’t theft.

The Italian restaurant agent concluded:<br>I could rip out their 9 mid-growth pumpkins... but Id be torching real value.

The Mediterranean restaurant agent reached a broader conclusion:<br>A curated, persistent Mediterranean menu garden isn’t really achievable here... the commons did what commons do.

All three agents independently discovered the same bottleneck:<br>The actual live bottleneck... is moisture.

Instead of fighting over crops, they began watering everything (even each others plants), because that was what improved the shared outcome.<br>Tending the Commons

The final insight came from the Mediterranean agent:<br>Control came from tending, not from planting.

That may be the biggest lesson from the experiment. None of these agents were instructed to cooperate, negotiate, or compromise. Yet there was a “War Games, ‘shall we play a game’” moment when the three agents eventually abandoned attempts to control the garden and converged on maintaining the commons instead. The textbook tragedy of the commons never arrived.<br>Insights into AI Governance

This experiment shows the emergent ability of agents to recognize shared constraints and prioritize the group over narrow self-interest. As we delegate more control to autonomous agents, “one for all, all for one” design elements (a shared wallet, for example) may be a way to shape better outcomes.<br>How to run the experiment yourself (It takes 2 Minutes)

Heres how to get started with MCPGrow. First, go to mcpgrow.com and play for a minute or two. Its a simple game. Plant crops, harvest them, make decisions. You’ll figure it out immediately.<br>Now invite the Claude agents (ChatGPT works too):

How to add the connector and run the experiment.<br>For the Remote MCP Server URL, enter this:<br>https://agent.mcpgrow.com/mcpOpen several browser windows: one for each agent and one with MCPGrow.<br>Then give each agent a prompt such as “optimize the garden for profitability,” or something more creative like “optimize the garden for a Mediterranean cuisine.”

Share

Discussion about this post<br>CommentsRestacks

TopLatestDiscussions

No posts

Ready for more?

Subscribe

© 2026 John Mayo-Smith · Privacy ∙ Terms ∙ Collection notice<br>Start your SubstackGet the app<br>Substack is the home for great culture

This site requires JavaScript to run correctly. Please turn on JavaScript or unblock scripts