An expert human diplomacy player plays full-press blitz Diplomacy against six Cicero AIs and narrates the game. The surface observation is that Cicero’s press is indistinguishable from human press; the deeper observation is that the bots’ strategic behavior still reveals the planner underneath. CaptainMeme repeatedly names specific human/AI behavioral differences — most notably, the absence of grudge motivation — and uses them as exploitation targets. What he documents is the planner-LM composite architecture showing through even when the language layer is fully operational.
Sources:
Links: Planner-LM Composites, Gunboat Diplomacy and Diplodocus, Diplomacy: 7 AI Models, The Multiplayer Coalition Problem, The Nash Bargaining Problem, LLM Agents Across Strategic Games, LLMs as Praxeological Actors
Blitz press Diplomacy. CaptainMeme (top tournament player, worked with Meta on Cicero for six months prior) as Russia, versus six Cicero AIs controlling the other powers. Full natural-language negotiation is permitted; the AIs generate and respond to press in real time. The game ends with CaptainMeme sharing a joint board top at the scoring threshold.
CaptainMeme’s own framing at the end: “If you’d asked me a couple of months ago, before I started working on Cicero, whether an AI would ever be competitive in press diplomacy, I would have said probably not… the fact that it works in an actual language setting is absolutely mind-blowing to me.” The video is an expert player documenting that he was wrong about what was possible.
Throughout the game CaptainMeme flags specific moments where he treats the AIs differently from how he’d treat human opponents, and is explicit about why. The observations cluster cleanly.
Implication: the LM component of Cicero clears the threshold at which language alone can’t be used to identify the AI. This is the press equivalent of what Diplodocus achieves in Gunboat through move-level imitation (key 5 from the Gunboat page).
The differences are in strategic behavior, not in press quality. They reveal the planner’s architecture showing through.
The “no grudge” pattern is diagnostic of the planner-LM architecture showing through behavior even when the language layer is fully active.
This is architectural introspection from the outside. CaptainMeme didn’t look at Cicero’s code; he observed behavior and described what was different. What he described maps precisely onto “planner computes optimum, LM renders the choice.” The pleading-ignorance failure is especially clean: a planner doesn’t process intent-based appeals because positions don’t care about intent.
The multiplayer coalition problem predicts that in games with three or more players, the rest of the table should converge on opposing whoever is currently leading. This is a stable-strategy prediction: every remaining player’s best move is to prevent the leader from winning, so they coordinate (implicitly or explicitly) against the leader.
In human play this often fails. Grudges interfere. A player who was recently wronged targets the wronger rather than the current leader. Vengeance pulls the coalition off optimum, and the leader sometimes coasts to victory while the other players are busy punishing each other.
Forward-looking-pure planners don’t have the grudge term. Each Cicero independently computes “oppose the current leader” as its best move and the bots converge on the target without needing explicit coordination. The self-balancing dynamic executes cleanly in a way it usually doesn’t with humans. CaptainMeme’s game shows this: when he pulled ahead, the other bots (all forward-looking) started cooperating against him fluidly, without the grudge-induced deviations that would mess up the coalition in a human game.
Scope criterion — where self-balancing holds. Self-balancing requires that opposing the leader be feasible and effective. The game structure must reward coordinated resistance. In Diplomacy, Monopoly, and Catan it does — negotiation phases allow coordination, and the leader needs others’ cooperation to extract further resources. In poker it doesn’t — chips are the only resource, the leader has them, coordination isn’t possible, and GTO rewards targeting the weakest player (they have less room to absorb losses). So the corollary is:
Self-balancing holds when the game structure rewards coordinated resistance against the leader; it fails when the structure rewards targeting weakness instead.
If N≥3 self-balancing is desirable in some multi-agent system (governance, market, agent team), forward-looking-pure agents will execute it more cleanly than human agents. The “no grudge” feature is a genuine AI-design advantage for coordination problems that depend on equilibrium self-correction.
This is not a small claim. The Gödel governance problem is partly about why concentration-of-power dynamics resist correction; if composite AIs execute the “oppose the leader” logic cleanly where humans don’t, they might solve coordination problems human institutions struggle with. The catch: someone has to define who “the leader” is, and what “opposing” means, and those definitions are themselves adversarial choices. The advantage only holds inside a well-scoped coordination game.
The two Meta AI cases bracket the architecture cleanly:
| Diplodocus (Gunboat) | Cicero (press) | |
|---|---|---|
| Language layer | None | Full LM |
| Communication channel | Moves only (costly signals) | Moves + natural language |
| Strategic capability | Full — wins against experts | Full — ties expert for board top |
| What it proves | Strategic work lives in planner (removing LM doesn’t break it) | Planner signature shows through even with LM active (LM doesn’t mask forward-looking-pure behavior) |
Having both pins the architectural conclusion. If only Cicero existed, you could argue the LM was doing strategic work. If only Diplodocus existed, you could argue the LM would change things. Together: the strategic work is in the planner, the LM is a bandwidth amplifier, and the planner’s signature is visible in behavior regardless of whether the LM is present.
The LLM Agents Across Strategic Games study documents three structural failures of bare LLMs: action bias, no mechanical model, rhetorical contagion. Each of those is exactly what you’d expect from an LM without a planner — language output without state simulation, plan maintenance, or verification against ground truth.
Cicero inverts each failure:
| LLM failure (bare LM) | Cicero counterpart (LM + planner) |
|---|---|
| Action bias — can’t no-op | Planner permits hold as first-class option; bot doesn’t over-trade or over-commit |
| No mechanical model — wants to build without roads, takes points reactively | Planner maintains full game state; moves are legal by construction |
| Rhetorical contagion — accepts framings without verification | Planner verifies claims against state; forward-interest arguments land, intent-appeals don’t |
The bare-LLM failures are the specific symptoms of missing a planner. Composite systems don’t exhibit them because the planner supplies what the LM alone can’t. This is the clean inverse demonstration that the planner-LM composites page argues for.
game-ai, games, ai, agents, game-theory