tl;dr: Smart agents fail at coordination the same way humans do. Better reasoning doesn't fix it. Better interaction does. Intelligence is not just in the agent. It is in the interaction.

- Ayush Chopra, PhD Candidate at MIT (media.mit.edu/~ayushc).
Learn more at - iceberg.mit.edu/#writing

For over thirty years, MIT's John Sterman ran a game called Fishbanks with human teams [1]. Six fishing companies share the same ocean. Each round, every company decides how many boats to send out. More boats means more profit — but the fish regenerate slowly. If too many companies fish too aggressively, the population collapses.

Round after round, teams overfished early, told themselves they'd pull back next turn, and watched others do the same. By the time anyone acted, the fish were gone. Populations crashed within 5–6 rounds and never recovered. The problem was never intelligence or intent. It was coordination, the inability to stop and restart together at the right moment.

We ran the same experiment with AI agents. Six identical agents, all greedy, all maximizing their catch over 35 rounds. The only variable we changed was how coordination was handled.

In one setting, coordination lived inside the agents. Agents shared their decisions and reasoning traces with neighbors but had no structured way to aggregate decision flexibility. Each independently inferred where the group was heading. This is the dominant approach in multi-agent systems today: decision-first coordination, where social inference remains part of individual cognition. We use the term "A2A" to refer to this style of coordination, agents exchanging decisions and reasoning, as supported by protocols like Google's A2A [2].

In the other setting, each agent ran a lightweight coordination layer locally called the Ripple Effect Protocol (REP) [3]. Agents still made their own decisions, but coordination logic was handled by a protocol layer running alongside them. There was no central controller and no shared objective, only a different way for actions to interact across the group.

Same agents. Same greed. Different coordination.

Part 1: The Outcome

Both systems collapsed early. In the first few rounds, agents in both settings overfished aggressively, driving the fish population from roughly 2,500 to under 200. This is not a failure of coordination — it is the natural result of greedy agents facing delayed feedback. No protocol eliminates that first mistake. The difference is what happens next.

Under A2A, the population never stabilized. Most agents eventually reduced fishing, but not at the same time. One agent would hold back while another surged. During fragile recovery windows, even a single defection was enough to push the system back toward collapse. Restraint without synchronization is not restraint at all. The same pattern observed in decades of human Fishbanks experiments reappeared: collapse without recovery.

REP agents crashed the fish population just as hard. But when the environment began showing signs of recovery, fishing pressure dropped across agents at roughly the same time. Not because they agreed to cooperate, but because the protocol layer propagated adaptation signals that shifted each agent's behavior before it acted again. The population rebuilt. When agents later overshot and triggered another decline, they slowed down together and recovered again. The pattern was not one recovery but repeated, coordinated recovery cycles (Figure 1).

Takeaway 1 — Coordination determines whether systems recover from a crisis. Both systems crashed early, driving the population below 200. The difference was recovery. A2A agents remained stuck at 29. REP agents coordinated systematically to rebound the fish population to 701 — making the environment 20x more sustainable (Figure 1).

Takeaway 2 — A2A agents fail the same way humans do. REP agents break the pattern. A2A agents reproduced the same collapse dynamics observed in decades of human Fishbanks experiments. Changing only the coordination layer, without altering incentives or decision-making logic, reversed the outcome. The difference was not smarter agents but better interaction.

Takeaway 3 — Coordination is not altruistic. It benefits everyone. REP agents caught 4,480 fish compared to 3,454 under A2A. Cumulative profit reached $44,100 under REP versus $25,324 under A2A (Figure 2). REP agents did not sacrifice competitiveness — they earned more because the shared resource remained viable.

Part 2: Why This Happens - Coordination is a Protocol Primitive

Under A2A, each agent has to do two jobs at once: decide what to do, and infer where the group is heading. As scale increases or feedback is delayed, that inference becomes fragile. An agent might correctly read the situation (fish stocks are low, restraint is needed) and still act at the wrong time, because it has no way of knowing whether others are about to restrain or surge. The bottleneck is not intelligence. It is where coordination sits.

REP separates these concerns. Agents still make their own decisions. But alongside each decision, they share a lightweight signal called a sensitivity, a statement describing how their behavior would change if conditions shifted. In Fishbanks, a sensitivity might read: "Fish population shows recovery signs based on catch composition. Fleet maintenance costs justify a conservative approach if others also coordinate." These are not commitments or promises. They are conditional signals, expressions of flexibility.

The protocol layer collects sensitivities from neighbors, aggregates them, and updates local coordination variables that shape the agent's next decision. No agent sees the full picture. No central controller exists. But as sensitivities ripple through the network, each agent's behavior begins to reflect not just what others did, but how the group is poised to respond. Agents don't react to the last round. They adapt to the emerging direction of the group before acting again.

Under A2A, the turning point for recovery must be inferred independently by each agent, often arriving too early or too late. Under REP, adaptation signals propagate before the next decision is made. The improvement is not smarter reasoning. It is better timing.

Intuition: Biology offers a useful parallel. Cells do not coordinate by announcing plans. When stressed, they emit chemical signals that neighboring cells respond to automatically. Local damage becomes a collective condition because adaptive signals propagate across the tissue. REP follows a similar principle. Agents remain autonomous, but adaptation signals flow through a decentralized protocol layer before decisions are taken again. Coordination emerges from how those signals interact, not from a central brain or shared objective.

The core insight is simple: cognition and coordination are different problems. Agents decide. Protocols shape how those decisions combine. When coordination is treated as infrastructure rather than inference, collective behavior changes — even when individual intelligence does not.

Part 3:  Beyond Symmetry - Coordination in the Beer Game

Fishbanks is symmetric. Every agent plays the same role in the same environment. Real coordination problems are rarely that clean.

We tested REP on the Beer Game, a classic supply-chain experiment involving four asymmetric roles—retailer, distributor, supplier, and factory.  Each sees only local inventory and orders from their immediate neighbor. No one has a global view. A small bump in customer demand at the retail end shouldn't cause problems, but it does. Each player, facing a shipping delay, slightly over-orders to be safe. That over-order becomes the next player's demand signal. By the time it reaches the factory, a 10% increase in customer demand has become a 200% spike in production orders. This is the bullwhip effect. Human teams, including trained MBA students, have struggled with it for decades.

Under A2A, AI agents reproduced the same oscillatory pattern. Small demand shocks amplified into costly swings, mirroring the boom-and-bust behavior documented in human experiments.

Under REP, those oscillations dampened. Orders stabilized more quickly, inventory swings narrowed, and total supply-chain cost decreased. Iceberg’s REP outperformed both Google’s A2A agents and Sloan MBA teams operating under the same constraints.

The Iceberg REP paper studies these dynamics under a broader range of scenarios, including adversarial behavior, heterogeneous network topologies, and privacy constraints. Further, AgentTorch framework [4] provides an SDK for deploying and scaling the REP across large populations of agents.

But how good can coordination get? To answer that, we computed a theoretical benchmark in our setup: the cost achievable under centralized coordination with full system information, $2619 in our setup (Figure 3). A controller with complete visibility and aligned incentives can dampen oscillations further and achieve lower total cost than either decentralized approach. That benchmark represents a north star—perfect coordination under ideal conditions.

REP narrows the gap to that benchmark without central control. But it does not eliminate it.

Part 4: What’s Next - Scaling Coordination

 Fishbanks involved six actors. The Beer Game involved four. The systems we are building, from supply chains and energy grids to financial markets and labor platforms, involve millions. At that scale, coordination failures don't stay local. They cascade. A thousand drivers responding to the same surge signal. A supply chain amplifying a rumor into a shortage. Individually rational decisions producing collectively irrational outcomes.

Project Iceberg designs and measures coordination at this scale. Can humans and AI agents coordinate at the scale of the systems they now inhabit?

If intelligence lives in the interaction, then the answer will not come from building smarter agents. It will come from building better coordination. 

How? Our breakthrough in Large Population Models is making this a reality!

Read more about our research and breakthroughs at iceberg.mit.edu/#writing

References:

[1]: mitsloan.mit.edu/teaching-resources-library/fishbanks-a-renewable-resource-management-simulation

[2]: github.com/a2aproject/a2a

[3]: The Ripple Effect Protocol: iceberg.mit.edu/protocol 

[4]: mitsloan.mit.edu/teaching-resources-library/mit-sloan-beer-game-online

[5]: github.com/AgentTorch/rep