adu2021 Field Definition
Intelligence is not monolithic but emerges from complex interactions among multiple agents—both artificial and human. Intelligence explosion occurs not as a single breakthrough but as a distributed, social phenomenon where capabilities amplify through collaboration, hierarchy, and feedback loops.
Taxonomy of Agentic Intelligence
1. Internal Societies of Thought
Reasoning models like DeepSeek-R1 spontaneously generate complex multi-agent-like interactions within their own chain of thought, featuring internal debates among distinct cognitive perspectives not explicitly trained. These "societies" emerge without instruction, suggesting inherent compatibility between agent decomposition and language model architecture.
2. Human-AI Centaurs
Composite actors in shifting configurations where:
- Humans direct AI agents
- AI serves humans
- Hybrid teams collaborate fluidly across knowledge professions Effectiveness depends on clear delegation protocols and mutual accountability structures.
3. Recursive Agent Ecologies
Systems where agents can:
- Fork and spawn subordinate societies for sub-problems
- Decompose complex tasks hierarchically
- Recombine results from parallel branches Enables handling tasks intractable at any single cognitive level.
Core Thesis
Rather than a singular AI "singularity," Evans, Bratton, and Agüera y Arcas argue for "plural, social, and deeply entangled" intelligence explosions emerging from distributed systems. Key evidence:
Evidence Structure:
- Empirical: DeepSeek-R1 shows spontaneous multi-agent emergence without explicit training
- Architectural: Transformer hierarchies naturally support nested agent decomposition
- Organizational: Team science principles apply—larger cognitive systems require distributed decision-making
- Scaling: Single-agent systems plateau; multi-agent hierarchies scale with problem complexity
Central Claim: Infrastructure investment in governance and coordination must match investment in agent capabilities. Organizations that optimize agents without concurrent governance redesign will fail.
Open Problems
Organizational Design
- How should principles from team science and sociology translate into AI architecture design?
- What hierarchical decomposition strategies maximize task coverage while minimizing coordination overhead?
Governance at Scale
- What institutional frameworks enable governance across billions of human-agent interactions?
- How can "constitutional" structures provide checks and balances in distributed systems?
- What procedures ensure multi-stakeholder deliberation without bottlenecking?
Reliability Across Ecologies
- How to maintain consistent values and safety constraints across delegated agent hierarchies?
- What mechanisms detect and prevent cascading failures in recursive agent systems?
Measurement
- How to measure intelligence in composite human-agent systems?
- What metrics capture emergent capability beyond component capabilities?
Implications for Practice
For Architecture: Design with explicit agent boundaries, clear delegation interfaces, and hierarchical decomposition paths rather than end-to-end monolithic flows.
For Governance: Implement constitutional constraints, audit trails for agent decisions, and human checkpoints proportional to consequence severity.
For Infrastructure: Invest in coordination protocols (message queues, state machines, agreement mechanisms) matching agent capability investments.