- Meta’s CICERO finished in the top 10% of Diplomacy players on webDiplomacy.net across 40 speed games, with human players frequently choosing to ally with it over other humans, the clearest public test of AI open-ended negotiation to date.
- In an independent 2026 study, 1,000 instances of the now-superseded Claude 3.5 Sonnet converged on a shared answer to a simple repeated choice without being told to, the largest group size tested, though the researchers caution this reflects statistical convergence rather than genuine coordination or social intelligence.
- A 2026 Deloitte report finds only 5% of organisations highly prepared for AI agents and just 15% at scaled adoption, with workforce readiness cited as a primary blocker.
Meta’s CICERO scored more than double the average human opponent in Diplomacy, anonymously, across 40 speed games, against players who often preferred working with it over other humans. That result, from November 2022, still stands as the clearest public proof that an AI agent can handle open-ended negotiation and persuasion at a competitive level. A separate 2026 academic study found that agents built on the now-superseded Claude 3.5 Sonnet converged on shared answers at a group size of 1,000, the largest tested, raising a different question: not whether agents can negotiate, but how far simple consensus dynamics scale before something more like coordination is needed.
CICERO’s Diplomacy Win
Diplomacy is a harder problem than Chess or Go. There are no deterministic paths to victory, just seven players forming alliances, breaking them and negotiating in free-form natural language throughout. Meta‘s CICERO played anonymously on webDiplomacy.net, finished in the top 10% of players who had competed in more than one game, and was rarely suspected of being an AI. Human players frequently chose to ally with it over other human participants.
The architecture paired a controllable dialogue model with a strategic reasoning engine. At each turn, CICERO read the board state and full conversation history, modelled likely opponent behaviour, then generated dialogue aligned with its own long-term plan rather than simply producing plausible-sounding responses. Meta acknowledged the system wasn’t clean, it sometimes contradicted its own stated objectives or produced inconsistent dialogue, but the overall performance cleared a bar most researchers had not expected an AI to reach this soon.
Reasoning Before Generation
The reasoning layer used an iterative planning algorithm Meta called piKL, which refined predictions about opponent moves and translated them into strategic intent before the language model produced any output. That sequencing is the key design decision. Most LLM deployments generate text in response to a prompt; CICERO generated text in service of a plan. The dialogue model accepted specific intents as input, so the system could offer support on one front while positioning for an alliance it would need three turns later.
Builders working on LangChain-based or LlamaIndex pipelines will recognise the pattern: separate the reasoning pass from the generation pass, and let the former constrain the latter. CICERO applied that at a level of social complexity, real-time persuasion across seven opponents, that most agentic frameworks haven’t approached. For a look at how recent updates to CrewAI and LlamaIndex are shifting agentic research costs the architecture decisions map directly onto what CICERO demonstrated.
Outside the Board Game
Meta’s researchers framed CICERO’s Diplomacy performance as a stepping stone toward agents that can manage long-running conversations, teach skills or handle real-time negotiations in sales, M&A or legal contexts. The gap between that vision and current capability is real. CICERO’s models are custom-built for Diplomacy’s specific structure, its interpretive and persuasive abilities don’t transfer cleanly to domains with different rules, different stakes and no clearly defined win condition.
The balance between strategic reasoning and natural language generation that makes it strong inside the game is also what makes it brittle outside it. Each new domain requires significant re-engineering of both layers. How widely applicable the underlying architecture actually is remains an open question; the available evidence is limited to the one setting where it has been tested at scale.
Enterprise Adoption Stuck in Pilot
According to a 2026 Deloitte report, only 5% of organisations say their business processes are highly prepared for AI agents, and just 15% have reached scaled, cross-functional multi-agent adoption, with workforce readiness and process investment ranking as the primary barriers. The gap between what CICERO and the 1,000-agent consensus study demonstrate in research settings and what enterprises can actually deploy is wide.
On the talent side, organisations lack the specialised skills to design, deploy and manage agentic systems, expertise that goes beyond traditional software development into prompt engineering for multi-agent interactions, AI ethics and integration with existing infrastructure. Upskilling takes time and money, and most enterprises haven’t started in earnest. Data is the other blocker. Agentic systems built for strategic reasoning need clean, well-structured datasets; enterprise data is frequently fragmented, inconsistent and incomplete. Transforming those disparate sources into something usable for training and operating agents is significant data engineering work, not a one-time fix. CICERO’s brittleness outside Diplomacy is a useful illustration of what happens when that re-engineering doesn’t happen, the underlying architecture simply doesn’t generalise. Questions about how human behaviour undermines AI governance at the implementation stage sit directly in this gap.
The 1,000-Agent Coordination Problem
A 2026 study’s finding that 1,000 agents running on the now-superseded Claude 3.5 Sonnet can converge on a shared answer without instruction points to a different problem class than CICERO’s one-agent-persuades-humans setup, though it’s a much simpler phenomenon: statistical consensus on a repeated binary choice, tested on a model since surpassed by several generations. Scaling coordination across agents that each have their own context window, their own partial view of a task and no shared memory is an orchestration challenge current frameworks handle poorly. AutoGen and CrewAI are pushing toward persistent memory, proactive problem detection and multi-step planning that goes well beyond the one-to-two step chains most production agents run today.
Whether that produces something closer to CICERO’s social fluency, or just faster parallel task execution, depends on how the reasoning and communication layers are designed. Those remain largely unsolved, which is why the hardware side of strategic agent development is drawing as much attention as the software.
