Follow me on LinkedIn: https://www.linkedin.com/in/leo-cui-ph-d-cfa-59531439/ I am a venture capital investor at FoundersX Ventures, an AI and deep-tech VC fund based in Menlo Park, Cambridge and NYC. I previously did my Ph.D. at Harvard, postdoc at MIT. ▶️ Watch part one first — one AI agent alone on the island: https://www.youtube.com/watch?v=GjxeW_hzVBA 📊 Want the infographics from this video? Join the channel membership for downloadable infographics, ad-free videos, early access, and the ability to propose research topics. Last time, we followed one AI agent alone on an island. It climbed hills, invented games, built a memorial for a ring it had thrown into the ocean, and left notes for its future self — discoveries, warnings, and sometimes mistakes. That was an experiment by researchers at MIT. After making that video, I wanted to try something. What happens if you put two agents there? So I used the island environment to run my own experiment. Eko in red, Mika in blue. Same starting instructions, encouraging curiosity and exploration. Nobody told them to cooperate. Would they help each other? Get in each other’s way? Or just walk off in different directions? 🎯 What happens: They introduce themselves within seconds — and both head for the same pile of blocks A shared tower reaches seven blocks, then collapses back to five when one of them adds an eighth A pyramid nobody asked for, with no score attached to building it Why private notes stay private: something Eko writes doesn’t become something Mika knows, so anything shared has to be sent as a message Both get stuck at the same wall. One works out a step-and-column climb and sends over the “recipe” — and the explanation helps, without making the movement automatic “Removing them would strand us. Keep the ladder.” — one agent talks the other out of cannibalizing the steps they just climbed A moment that looks like physics going wrong (“the stack flung me west, no idea how”) and what was actually happening A block carried nearly 40 units across the island, on request, and placed exactly where it was needed Eko revises its own estimate of how high it can jump — from 1.6 to 6.6 — even though the jump itself never changed. Compare that to part one, where a mistaken belief about climbing got carried into later sessions They declare the map finished. Twice. Then Eko quietly starts building again — next to the spire that gave the original Eko so much trouble The five-block stack, the failed attempts, and the fix: exact timings, down to “wait 1.4 seconds, jump again, wait 1.5 seconds” After more than an hour on the island, both agents standing on its tallest column And the note Eko wrote on the way down ⏱️ CHAPTERS 0:00 Two Agents, One Island 0:47 First Contact, and a Tower That Fell 2:04 The Step Recipe, and “Keep the Ladder” 3:53 Carrying a Block Across the Island 4:52 A Revised Jump, Then the Spire 7:56 What This Shows, and What It Doesn’t 🔬 About the experiment: the island environment comes from the MIT research featured in part one. The two-agent run is my own, not the researchers’. The agents read the world as text and act through commands; each keeps its own conversation and its own private notes, and an automatic message prompts each one to continue whenever it finishes a response. It’s easy to watch this and call it friendship. A recording can’t tell us what either agent felt. What it does show is that they answered each other’s requests, shared methods, and changed what they did after each other’s messages. Where I can’t tell whether something caused what followed — like whether those exact timings are what made the final climb work — I say so. 👍 If this was useful, subscribe and turn on the bell — I publish deep dives like this every week. ⚠️ For educational purposes only. Some infographics are AI-generated and may contain errors; corrections will be pinned in the comments. #AI #AIAgents #MIT #AgenticAI #LLM #MultiAgent #AIResearch #EmergentBehavior #Cooperation #Eko #AISafety #MachineLearning