Why AI researchers are ditching mega-models for Minsky’s multi-agent approach

Marvin Minsky’s 1986 book “The Society of Mind” is finding new relevance in 2025 as AI researchers increasingly embrace modular, multi-agent approaches over monolithic large language models. The theory, which proposes that intelligence emerges from collections of simple “agents” rather than a single unified system, now maps directly onto current AI architectures like Mixture-of-Experts models and multi-agent frameworks such as HuggingGPT and AutoGen.

Why this matters: As the AI field hits the limits of scaling single massive models, Minsky’s vision offers a blueprint for building more robust, scalable, and aligned AI systems through modularity and internal oversight mechanisms.

The core theory: Minsky argued that “the power of intelligence stems from our vast diversity, not from any single, perfect principle.”

• The mind consists of countless simple agents that individually do little but collectively produce complex thinking.
• These agents form hierarchies where higher-level agents coordinate lower-level ones.
• Special “censor” and “suppressor” agents provide internal oversight to prevent dangerous or unproductive behaviors.
• A “B-brain” monitors the primary “A-brain” processes, watching for errors and intervening when necessary.

Current limitations of monolithic AI: Today’s large language models face significant constraints despite their impressive capabilities.

• Single models struggle with multi-step reasoning, long-horizon planning, and lack built-in mechanisms to check their outputs.
• They can “hallucinate false information with supreme confidence” and don’t know when they’re wrong.
• Having one model handle complex multi-faceted tasks often leads to loss of coherence or errors.
• The “one model to rule them all” approach shows diminishing returns as scaling becomes less effective.

Mixture-of-Experts as modern implementation: MoE architectures embody Minsky’s modularity principles in practice.

• These models split neural networks into specialized sub-networks (experts) with a gating mechanism routing inputs to appropriate experts.
• Only a few experts activate for any given input, making computation efficient while enabling trillion-parameter models.
• Each expert develops distinct skills, similar to agents in Minsky’s society with different roles.

Multi-agent systems in action: Frameworks like HuggingGPT and AutoGen are creating literal AI societies.

• HuggingGPT uses a large language model as controller to manage other specialized models for complex tasks.
• AutoGen enables multiple LLM agents to converse and collaborate, with customizable roles like brainstormer and critic.
• These systems often adopt functional roles reminiscent of Minsky’s agencies: planners, workers, critics, and memory stores.
• Multi-agent approaches excel at tasks requiring decomposition and iteration, operating in parallel for speed and scalability.

AI alignment through internal critics: Minsky’s censor agents and B-brain concept directly inform current alignment research.

• Self-reflection techniques where LLMs critique their own answers significantly improve correctness.
• “LLMs are able to reflect upon their own chain-of-thought and produce guidance that can significantly improve problem-solving performance.”
• Multi-agent debate systems use adversarial dialogue between AI agents to surface truth more effectively.
• Constitutional AI and similar approaches implement internal oversight mechanisms to catch harmful outputs.

The centralized vs. decentralized debate: The AI field is experiencing a pendulum swing toward modular architectures.

• Monolithic systems offer simplicity and potential emergent properties from integrated training.
• Modular systems provide flexibility, specialization, and fault tolerance with component-level optimization.
• Current trend favors “coordination over raw scale” as practitioners build societies of models rather than single mega-models.
• Hybrid approaches may prove optimal, using large models as components within larger orchestrated systems.

What practitioners are saying: The AI community increasingly recognizes the value of Minsky’s approach.

• “Multi-agent setups today are basically operationalizing Society of Mind… it’s coordination over raw scale now,” noted one AI engineer.
• Developers report that “a solver proposes, a critic flags issues, and a refiner improves the answer – this kind of structure consistently improves factual accuracy.”
• However, some warn that modularity can introduce new failure modes at interfaces and potentially reduce emergent behaviors.