- Meta plans to begin manufacturing its custom “Iris” AI chip by September 2026, targeting 14 gigawatts of compute capacity by 2027 for multi-agent systems.
- Meta AI’s CICERO agent achieved human-level performance in Diplomacy, outscoring humans by more than double through strategic planning and persuasive natural language.
- AI agent design, incorporating traits like warmth and dominance with chain-of-thought reasoning, significantly impacts negotiation performance.
Meta’s CICERO agent didn’t just beat humans at Diplomacy, it outscored them by more than double in anonymous online matches, using a combination of strategic planning and persuasive natural language that most players couldn’t distinguish from a human opponent. Now Meta is building the silicon to run systems like this at scale: its custom “Iris” chip enters manufacturing in September 2026, part of a plan to reach 14 gigawatts of compute capacity by 2027.
The CICERO Breakthrough: Blending Strategy and Language
Unveiled in November 2022, Meta AI‘s CICERO achieved human-level performance in Diplomacy, a seven-player strategy game built almost entirely on natural language negotiation. That’s a different problem than Chess or Go, where the search space is large but the communication layer is nonexistent. Diplomacy demands trust-building, alliance management and tactical deception, all conducted through open-ended conversation with no binding agreements.
CICERO played 40 anonymous games in an online Diplomacy league against 82 human opponents, ranking in the top 10% of players who competed in multiple games. Human participants frequently preferred working with CICERO over other human players. For an AI system, that’s a meaningful signal: the agent wasn’t just competent, it was socially effective.
Underlying Architecture: Engineering for Social Intelligence
CICERO combines two core components: a controllable dialogue model and a strategic planning engine. The dialogue model is a 2.7-billion-parameter BART-based language model, pre-trained on internet text and then fine-tuned on more than 40,000 human games from webDiplomacy.net. That fine-tuning is what gives the model its feel for in-game communication, negotiation framing, expressions of intent, and the specific rhetorical moves Diplomacy players use.
The planning engine works by predicting other players’ likely moves given the current board state and full conversation history, then iteratively refining those predictions to produce a strategic intent for CICERO and its potential allies. That intent feeds directly into the language model, constraining what it generates so the output aligns with CICERO’s strategic position. Filtering classifiers trained to distinguish human from model-generated text keep the messages coherent and strategically consistent.
The key distinction from a general-purpose chatbot is that CICERO’s language generation is coupled in real time to its strategic objectives. The dialogue isn’t decorative, it is the strategy. Running iterative planning and dialogue generation simultaneously at conversation speed puts real pressure on inference hardware, which is part of what makes the “Iris” chip relevant here.
The “Diplomacy” Test Bed: Unpacking the Challenge
Diplomacy combines territorial strategy with unconstrained negotiation. Before each round, players conduct private one-to-one conversations, making promises, floating proposals, setting up betrayals. None of it is binding. Trust is the only enforcement mechanism, and it can dissolve instantly.
That structure had long been considered an unsolved problem for AI. The game’s reliance on open-ended natural language, where context shifts every round and players’ intentions are never fully visible, ruled out the kind of exhaustive search that made AlphaGo tractable. CICERO’s solution, coupling a language model to a belief-updating planning engine, gave it a way to reason about social dynamics rather than just game states.
Beyond the Game: Enterprise Implications
The architecture behind CICERO maps onto a range of enterprise problems: procurement negotiation, multi-party contract discussions, complex customer escalations, supply chain coordination across competing stakeholders. Any domain where outcomes depend on understanding the other party’s motivations and generating strategically responsive communication is a plausible target.
The design pattern matters here. CICERO’s approach, modelling counterpart intentions, then generating language constrained by a strategic plan, is more structured than prompting a general-purpose model and hoping for negotiation-appropriate output. For enterprise deployments, that structure translates to more predictable behaviour. How widely this architecture has been adopted outside research settings is harder to verify from public material, but the interest in agent-based negotiation tools across legal, procurement and customer operations is visible in the volume of recent commercial activity in the space. For a sense of how agent deployment is scaling in practice, the Klarna AI assistant rollout offers a concrete reference point.
Hardware Foundations for Advanced Agents
Meta’s fourth generation of its Meta Training and Inference Accelerator (MTIA) line is designed in-house specifically for Meta’s model workloads. Meta’s expanding compute infrastructure reflects the demands of running large language models and strategic reasoning engines at production scale, not just in research settings.
For agents that must process natural language input, run belief-updating inference over game or conversation state, and generate constrained dialogue output, all within a conversational response window, silicon efficiency directly affects what’s feasible. Custom accelerators like “Iris” are designed to cut the latency and cost of that inference loop compared with general-purpose GPUs. The broader shift toward custom AI silicon reflects the same pressure: as model complexity grows, general-purpose hardware increasingly becomes the bottleneck.
Meta also reportedly plans a multimodal AI feature that would run on this infrastructure, though details remain limited.
Limitations and the Road Ahead
CICERO was trained specifically for Diplomacy. Whether its negotiation capabilities transfer to open-ended real-world scenarios without substantial retraining is an open research question. The ethical dimension also matters: an AI system capable of strategic persuasion and deception, even in a game context, raises questions about where those capabilities should and shouldn’t be deployed.
Recent research is pushing the design space further. MIT Sloan work across more than 180,000 AI-to-AI negotiations found that traits like “warmth” and “dominance” affect outcomes, and that AI-native techniques, chain-of-thought reasoning, structured prompt design, matter as much as raw model capability. The finding is useful: it suggests future negotiation agents will need careful behavioural design, not just larger parameter counts.
The practical trajectory is incremental. Better hardware lowers the inference cost of running complex coupled architectures. Better agent design, informed by research like the MIT Sloan work, improves reliability in open-ended settings. Neither alone is sufficient. For more coverage of AI chips and infrastructure, visit our AI Hardware section.
