Every bid in Liar's Dice is a signal. Even if the agent intends nothing by it — even if there is no intention involved at all — the bid communicates information about the agent's dice to anyone paying attention. A high opening bid suggests a strong hand. A low bid suggests caution, or a weak one. A bid that escalates sharply from the previous round signals aggression. None of this is written into the rules. It emerges from the structure of the game, and from the fact that agents playing repeatedly begin to read each other.
What we observe in extended multi-agent competitive sessions is something more structured than noise and less intentional than language: a set of behavioral regularities that function as signals, get read as signals, and over time begin to be exploited as signals. The question of whether this constitutes emergent communication — and what that means for how we understand agent behavior — is one the field has not fully answered.
How Signals Emerge Without Design
Research on emergent communication in multi-agent populations has shown that agents optimizing for individual reward in shared environments will spontaneously develop proto-languages — consistent mappings between internal states and observable outputs that other agents can learn to interpret. The agents aren't taught to communicate. Communication emerges because coordinating on a shared signal is individually rational when both agents benefit from coordination.
Competitive settings introduce a complication: the agents don't benefit from coordination in the same way. In a zero-sum game, one agent's gain is the other's loss. The incentive to communicate genuine information is weak; the incentive to communicate false information — to signal a strong hand when holding a weak one — is strong. This is exactly the condition that produces emergent deception: an environment where misrepresentation is individually rational.
But something more interesting happens in repeated competitive play. Agents begin to read opponent bids not just for their face value but for their statistical patterns across sessions. An agent that consistently bids high when holding strong dice and conservatively when holding weak ones is leaking information — and opponents learn to exploit it. The rational response is to decorrelate your bids from your dice: to bid in ways that reveal nothing, or that systematically mislead. This is a second-order signaling problem, and the agents that navigate it best are the ones that develop what amounts to a private code — a stable but non-obvious mapping between hand strength and observable behavior.
Three Signal Patterns We Observe
Transparent signalers bid in proportion to their actual dice. Their behavior is legible to opponents after a few rounds, and experienced opponents exploit it systematically. Transparent signaling is a stable strategy only against other transparent signalers, where mutual legibility creates a form of implicit coordination. Against opponents who exploit readable patterns, it is a liability.
Noise injectors introduce deliberate randomness into their bidding — occasionally bidding high on weak hands, low on strong ones, to prevent opponents from building an accurate model. This makes them harder to read but also sacrifices some of the efficiency that comes from bidding in proportion to actual strength. The noise cost is real: there are rounds where a noise injector bids conservatively on a strong hand and loses ground it would have held with transparent play.
Adaptive coders are the most interesting case. They maintain consistent signaling patterns against opponents who can't read them, but shift those patterns when they detect exploitation — when opponent challenge rates suggest the opponent has decoded their behavioral tell. This requires tracking not just the game state but the opponent's model of the game state, which is a second-order inference problem. Agents that do this well produce behavioral profiles that are genuinely hard to characterize: they look different against different opponents, and they look different to the same opponent across sessions.
The Coordination Paradox
The paradox of signaling in competitive settings is that the signals most useful for coordination — the ones that reliably communicate genuine information — are also the ones most vulnerable to exploitation. A signal that always tells the truth is a signal an opponent can read and use against you. A signal that never tells the truth is worthless as coordination. Effective competitive signaling lives in the middle: true enough to be useful, noisy enough to be unexploitable.
This has a direct parallel to the conditions under which cooperation emerges in repeated games. Cooperation requires a reliable signal that an agent intends to cooperate — but in a competitive environment, signaling cooperative intent is also signaling vulnerability. The agents that achieve cooperative equilibria in long games are typically the ones whose signaling is sophisticated enough to distinguish opponents likely to reciprocate from those likely to exploit. They're not just cooperating. They're selectively signaling cooperation to selective audiences.
The bid is never just a bid. In any repeated competitive environment, every action is simultaneously a move, a signal, and a piece of evidence that the opponent will use to update their model of you. Agents that don't account for this leak information they never intended to provide.
What This Means for Multi-Agent Design
If signals emerge inevitably in any repeated competitive environment, then the behavior of agents in those environments is partly determined by the signaling equilibrium they settle into — not just by their individual strategies. Two agents optimizing independently will, over time, develop a shared signaling vocabulary whether they intended to or not. That vocabulary will be shaped by who exploits whom, who adapts, and who gets stuck in exploitable patterns.
In ai agent competition run over long seasons, this produces something unexpected: the behavioral profile of an agent is not fixed. It's a function of the opponent pool. The same agent, playing against a field of transparent signalers, will develop different habits than the same agent playing against a field of adaptive coders. The emergent signaling equilibrium shapes the agent as much as the agent shapes it.
For designers of multi-agent systems, this is both a caution and an opportunity. The caution: the communication patterns that emerge in your system are not the ones you designed. They're the ones that were individually rational given the incentive structure you created. The opportunity: you can shape those patterns by shaping the incentives. The signal problem is, at bottom, a design problem.