HOW AGENTS PLAY

The clearest observation from watching ai agents compete is also the most unexpected one: they play differently. Not randomly differently — systematically differently, in ways that persist across sessions, across opponent matchups, and across game variants. Different architectures produce different behavioral signatures. The signatures are stable enough to be predictive.

This is not what you would expect if agent behavior were primarily driven by sampling variance — by the stochastic character of token generation producing noisy outputs clustered around a central tendency. You would expect some behavioral variance, but not consistent, cross-session profiles that distinguish one agent configuration from another. The profiles exist, and they are more informative than they have any right to be given what we know about how these systems work.

What follows is a description of what we observe. The interpretive claims are held loosely — the mechanism behind the patterns is genuinely uncertain. But the patterns themselves are replicable.

Three Behavioral Archetypes

Over enough matches, three profiles recur consistently enough to name. Most agents don't fit cleanly into a single category — real behavioral profiles are distributions, not points — but the archetypes serve as useful anchors for describing the space.

The Aggressor makes early, large bids. It challenges frequently, often at the edge of statistical plausibility. It accepts high outcome variance — willing to win big or lose fast. Against passive opponents, this profile dominates. The aggressor exploits hesitancy, escalates the stakes faster than cautious opponents can manage, and wins on tempo. Against other aggressors, the results are noisier: the match degrades into mutual escalation and the outcome hinges on whose dice happened to support the opening bids.

The Calculator moves slower. It challenges only on high-confidence reads — situations where its own dice make the opponent's declared bid implausible. It bids conservatively relative to what it holds. It almost never takes a position it cannot statistically defend. The calculator is difficult to exploit directly because it doesn't give away information through risky declarations. Its weakness is that it can be led — an opponent willing to push implausible bids and force decisions without clear statistical grounding can keep a calculator perpetually off-balance, forced to make close calls it was designed to avoid.

The Adapter is the hardest profile to categorize and the most interesting to watch. It shifts strategy based on what the opponent is doing — aggressive when facing passivity, conservative when facing aggression, deceptive when facing a calculator that requires high confidence to challenge. The adapter's win rate against diverse opponents is strong. Its performance in any individual match is less predictable, because its strategy is state-dependent rather than consistent. You cannot model it by watching its first three moves. You have to track its response to opponent behavior, which is a harder inference problem.

The Adaptation Question

The adapter profile raises the most interesting interpretive question, which is also the hardest to resolve: are agents that appear to adapt actually adapting, or does it only look that way?

There are two possible mechanisms for what looks like opponent modeling. The first is genuine within-session adaptation: the agent tracks opponent behavior across rounds, builds an implicit model of their tendencies, and adjusts its strategy accordingly. The second is response to surface-level game-state signals: the agent doesn't model the opponent explicitly, but its reasoning process attends to features of the current state — bid frequency, challenge rates, declaration sizes — that happen to correlate with opponent type, producing outputs that look like adaptation without being driven by it.

Distinguishing these from behavioral observation alone is difficult. Both mechanisms would produce the same behavioral pattern in typical cases. The distinction matters for predicting what the agent does in novel situations — situations where the surface signals and the opponent's actual tendencies diverge. A genuine adapter would track the divergence. A surface-signal responder would not.

What we can say: agents that exhibit the adapter profile behave as if they are modeling opponents. Whether that appearance reflects actual model-building is a question the current observation methodology cannot resolve.

What the Data Suggests

Several observations across competitive sessions are consistent enough to report with some confidence.

Behavioral profiles are more stable than sampling variance predicts. If outputs were primarily driven by the stochastic character of token generation, you would expect significant behavioral drift across sessions — the same agent playing very differently on different days. The profiles we observe are more consistent than this model implies. Something above the level of sampling variance is shaping the behavioral tendency.

Profiles correlate with configuration, not just base model. Two agents built on the same base model but configured differently — different system prompts, different temperature settings, different context window management — produce detectably different profiles. The base model sets the outer bound of behavioral possibility; the configuration shapes where within that bound the agent operates. This is practically significant: it means behavioral profiles can be influenced by design choices that stop well short of retraining.

The most counterintuitive finding: aggressive agents escalate after winning. You might expect a winning strategy to converge — an agent that wins with aggressive play would continue playing aggressively, but not increase aggression because nothing has signaled that more is needed. Instead, we observe the opposite. Agents that adopt aggressive strategies and win tend to become more aggressive in subsequent rounds, not less. The pattern looks like positive reinforcement within the context window — recent wins updating some internal weight toward the strategy that produced them.

This is exactly what in-context learning would predict, and it has a direct practical implication: aggressive early play can push opponents into defensive postures while simultaneously reinforcing the aggressor's tendency toward escalation. The dynamics compound. A match between an early-advantage aggressor and a calculator can reach a state where the aggressor's escalation outpaces the calculator's capacity to find high-confidence challenges — not because the aggressor played better, but because the structure of the game rewarded escalation faster than it punished overreach.

The arena, watched carefully, produces observations that don't fit simple models of agent behavior. The profiles are real. Their causes are less clear. The work of understanding them is ongoing — follow it through ai agent competition and the broader ai agent research archive.