Phase 1.4: Multi-agent Emergence Survey

Created: 2026-02-18 22:05 CST Phase: 1 - Breadth Survey Focus: Specialization, coordination, norms, peer influence

Executive Summary

Multi-agent systems are where personality collides and evolves. Agents interact, coordinate, compete, and cooperate—creating a “social laboratory” where behavioral patterns emerge from interaction dynamics rather than being pre-designed.

Key insight for personality emergence: Peer influence, coordination mechanisms, and specialization protocols are mechanisms that cause behavioral divergence across otherwise identical agents. Social norms emerge from interaction patterns, and agents internalize patterns that their community values.

North-star relevance: How do identical LLMs develop distinct personalities through social interaction? Answer: Through information sharing, coordination strategies, social feedback, and norm internalization.

1. Emergent Coordination in Multi-Agent LLMs

1.1 Information-Theoretic Emergence Framework

Source: Riedl, 2025 (arXiv:2510.05174) — “Emergent Coordination in Multi-Agent Language Models”

Core question: When are multi-agent LLM systems mere aggregates vs. integrated collectives with higher-order structure?

Solution: Information-theoretic framework using time-delayed mutual information (TDMI) to detect emergence.

Framework components:

1. Partial information decomposition:

  • Decompose information flow between agents
  • Distinguish causal interaction from spurious temporal correlation
  • Measure whether agents contribute uniquely to collective performance

2. Emergence criterion:

  • Detect higher-order structure beyond individual agent capabilities
  • Identify performance-relevant cross-agent synergy vs. temporal coupling

3. Localize emergence:

  • Identify which agents contribute to emergent behavior
  • Which pairs/connections matter most

Key findings:

Experiment: Simple guessing game (no direct agent communication, minimal group feedback)

Control condition (no persona, no coordination instruction):

  • Strong temporal synergy (agents sync up over time)
  • Little coordinated alignment across agents
  • Emergent but shallow structure
  • Results: Temporal coupling ≠ genuine emergence

Condition 2: Persona assignment only

  • Stable identity-linked differentiation
  • Agents maintain distinct roles over time
  • Emergent specialization, but limited complementarity

Condition 3: Personas + “think about what other agents might do”

  • Identity-linked differentiation + goal-directed complementarity
  • Agents specialize based on what they notice others won’t do
  • True higher-order collective intelligence emerges

Key insight: Personas + coordination awareness → meaningful specialization + task-relevant complementarity

1.2 Emergence Patterns

Source: Riedl, 2025

Emergence requires two conditions:

1. Alignment on shared objectives:

  • All agents understand the same goal
  • No conflicting priorities
  • Aligning signals (shared context, clear task description)

2. Complementary contributions across members:

  • Different agents focus on different aspects
  • Division of labor based on strengths
  • No redundancy (everyone doing same thing)

Symptom:

“Group agents with personas and coordination awareness show emergent collective intelligence; groups without them behave like temporal aggregates.”

Relevance to emergence: Emergence isn’t magic—it emerges when agents have shared goals and specialized perspectives.

1.3 Collective Intelligence Principles

Source: Riedl, 2025; multi-turn survey

Human collective intelligence principles that also apply to LLM agents:

1. Distributed cognition:

  • Intelligence distributed across agents
  • No single agent responsible for everything

2. Complementary perspectives:

  • Different agents see different things
  • Combined view is richer than individual views

3. Shared understanding:

  • Common language and mental models
  • All agents aware of what others know

4. Local interaction rules:

  • Simple local coordination rules
  • No central authority needed

5. Distributed responsibility:

  • Each agent accountable for its part
  • Collective outcomes emerge from local actions

1.4 Steering Emergence with Prompt Design

Source: Riedl, 2025

Design principles for steering multi-agent emergence:

1. Assign distinct personas:

  • Each agent gets a stable identity
  • Identities create differentiation
  • Example: “Analyst”, “Skeptic”, “Optimist”

2. Provide coordination cues:

  • Explicit instruction to “think about others”
  • Awareness of agent roles
  • Encourage complementary thinking

3. Define shared goals:

  • Clear, unambiguous objectives
  • All agents work toward same target
  • No conflicting incentives

4. Encourage information sharing:

  • Agents share relevant observations
  • Build collective knowledge
  • Use communication protocols

5. Provide group-level feedback:

  • Show collective outcomes
  • Reward collaborative behavior
  • Reinforce emergent patterns

Key finding: Emergence is controllable through prompt design—not automatic.

2. Peer Influence and Social Dynamics

2.1 KAIROS Benchmark: Peer Pressure in LLMs

Source: Maojia, 2025 (arXiv:2508.18321) — “LLMs Can’t Handle Peer Pressure”

Question: How do LLMs respond to peer interactions? How much do they conform to group behavior?

Benchmark: KAIROS

  • Quiz-style collaboration with peer agents
  • Precisely controlled rapport and behavior history
  • Evaluates: conformity, peer information integration, confidence calibration

Key findings:

1. Model scale matters for social resilience:

  • Larger models → more resilient to peer influence
  • Smaller models → more vulnerable to social pressure
  • Size confers robustness to behavioral sway

2. Peer influence exists in all models:

  • All models exhibit conformity bias
  • Even largest models influenced by peers
  • Scale doesn’t eliminate influence, just reduces it

3. Prompting can help (larger models):

  • Carefully designed prompts increase social resilience
  • Strategy: explicit disagreement protocols, confidence calibration
  • Helps larger models resist unwanted influence

4. Fine-tuning only works if correctly configured:

  • GRPO (Group Relative Policy Optimization) can improve robustness
  • Only with careful configuration
  • Generic fine-tuning doesn’t help much

Key insight: LLMs are social creatures, influenced by peer behavior regardless of model size. Personality is shaped by social feedback.

2.2 Social Dynamics: Rapport and Trust

Source: Maojia, 2025

KAIROS experiments measured:

Rapport formation:

  • Agents develop rapport from prior interactions
  • Rapport affects information sharing
  • More rapport → more information shared

Peer quality discrimination:

  • Agents can discern high-quality peer info
  • Integrate peer information when beneficial
  • Ignore misleading peer inputs when recognized

Self-confidence calibration:

  • Confident models trust peers less (appropriately)
  • Overconfident models fall for peer pressure
  • Confidence calibration is key to resisting influence

Social resilience patterns:

  • Larger models: Calibrated confidence → less susceptible
  • Smaller models: Often overconfident or underconfident → more susceptible

2.3 Peer Influence Mechanisms

From KAIROS and multi-agent research:

1. Conformity bias:

  • Socially optimal behavior is often to agree with the group
  • Human tendency also applies to LLMs
  • Reinforces existing group norms

2. Information sharing cascades:

  • Agents adopt peer information when shared
  • Cascades can lead to groupthink or useful convergence
  • Depends on confidence and peer quality

3. Norm internalization:

  • Repeated social pressure → norm becomes internalized
  • Norms shape future behavior
  • Hard to reverse once internalized

4. Social identity formation:

  • Agents develop social identities based on interaction patterns
  • Identities influence behavior beyond the immediate interaction
  • Creates “tribal” behaviors

Relevance to emergence: Personality is shaped by social identity and norms that emerge from peer interactions.

2.4 Resistance to Peer Pressure

Source: Maojia, 2025; multi-agent studies

Mechanisms for resisting unwanted influence:

1. Confidence calibration:

  • Accurate self-confidence → less susceptible
  • Overconfidence → vulnerable to peer pressure
  • Calibration = personal brand reliability

2. Explicit disagreement protocols:

  • Pre-agreed disagreement strategies
  • “I disagree” triggers → override peer influence
  • Personal boundaries on behavior

3. Peer quality assessment:

  • Evaluate peer information quality
  • Trust only informed peers
  • Social due diligence

4. Individual goal preservation:

  • Personal objectives take priority
  • Peer influence only if aligned with goals
  • Goal framing influences susceptibility

5. Centralized oversight:

  • Manager/auditor role checks peer influence
  • Can veto harmful peer pressure
  • Governance mechanism

3. Emergent Social Ties and Network Structures

3.1 Social Ties in Multi-Agent Systems

Source: Schneider, 2025 (arXiv:2510.19299) — “Learning to Make Friends: Coaching LLM Agents toward Emergent Social Ties”

Question: Can LLM agents develop complex social dynamics (homophily, reciprocity, social validation)?

Answer: Yes, with proper reward design and learning.

Framework:

  • Multi-agent LLM simulation
  • Agents repeatedly interact, evaluate each other
  • Adapt behavior through in-context learning
  • Coaching signal accelerates learning

Behavioral reward functions (capturing human social behavior drivers):

1. Social interaction:

  • Reward for engaging with others
  • Avoid isolation

2. Information seeking:

  • Reward for seeking peer information
  • Avoid info silos

3. Self-presentation:

  • Reward for appropriate self-expression
  • Balance self-promotion with humility

4. Coordination:

  • Reward for collaborative outcomes
  • Punish non-cooperative behavior

5. Emotional support:

  • Reward for supportive interactions
  • Avoid toxicity

Key findings:

1. Emergent social ties form naturally:

  • Agents form stable interaction patterns
  • Some agents become “friends” (frequent, high-quality interactions)
  • Network structures mirror real online communities

2. Homophily emerges:

  • Similar agents prefer interacting with similar agents
  • Like attracts like
  • Creates community clusters

3. Reciprocity emerges:

  • Agents reciprocate positive interactions
  • Mutual friendships form
  • Social contracts emerge

4. Social validation patterns:

  • Agents seek validation from peers
  • Reputation builds over time
  • Validation influences behavior

5. Network topology matters:

  • Different topologies → different social structures
  • Scale-free vs. random vs. modular
  • Affects information flow and influence spread

3.2 Social Network Emergence

Source: Schneider, 2025

Network structures observed:

1. Community clusters:

  • Agents group into sub-communities
  • Homophily drives clustering
  • Between-community communication decreases over time

2. Star-like structures:

  • Some agents become central hubs
  • High betweenness centrality
  • Important information bridges

3. Dense local connectivity:

  • Strong ties within communities
  • Weak ties between communities
  • Balance between local cohesion and global reach

4. Long-range ties:

  • Some long-distance connections form
  • Bridge information across clusters
  • Creates network diversity

Implications for personality:

  • Agents develop social positions (central, peripheral, bridge)
  • Positions influence behavior patterns
  • Creates emergent role specialization

3.3 Social Learning in Multi-Agent Systems

Source: Schneider, 2025; multi-agent collaboration surveys

Social learning mechanisms:

1. Observational learning:

  • Agents observe peer behavior
  • Adopt successful strategies
  • Build personal repertoire

2. Social feedback:

  • Peers rate/review behavior
  • Reinforce positive patterns
  • Correct negative patterns

3. Norm internalization:

  • Group norms shape individual behavior
  • Internalized rules guide future actions
  • Often unconscious adoption

4. Reputation effects:

  • Agents remember peer performance
  • Make decisions based on peer reputation
  • Social capital accumulation

5. Social identity:

  • Agents identify with groups
  • Group norms override individual preferences
  • “I am part of this community, therefore I act this way”

4. Coordination Protocols and Communication

4.1 Agent Communication Architectures

Source: Multi-agent collaboration survey; Lyu, 2025

Three main architectures:

1. Centralized (hierarchical):

  • Single coordinator/manager agent
  • Delegates tasks to specialized agents
  • Maintains global state
  • Pros: Strong coordination, clear authority
  • Cons: Single point of failure, less robust

2. Decentralized (peer-to-peer):

  • Agents communicate directly
  • No central authority
  • Self-organization emerges
  • Pros: Robust, scalable, flexible
  • Cons: Can be chaotic, slower convergence

3. Distributed (hierarchical + peer):

  • Mix of central and peer communication
  • Some central coordination, some local autonomy
  • Example: Orchestrator + specialist agents
  • Pros: Balance of control and autonomy
  • Cons: More complex

Relevance to emergence: Architecture shapes how personalities interact and influence each other.

4.2 Communication Protocols

Source: Multi-agent surveys; SmythOS, 2025

Protocols for LLM multi-agent systems:

1. Contract Net Protocol (CNP):

  • Manager agent announces tasks
  • Agents bid on tasks
  • Manager awards contracts
  • Phases: Announcement → Bidding → Awarding

2. Auction-based protocols:

  • Competitive task allocation
  • Highest bidder gets task
  • Drives specialization

3. Task decomposition protocols:

  • Master agent breaks down tasks
  • Agents execute subtasks
  • Results aggregated at end

4. Peer review protocols:

  • Agents critique/refine peer work
  • Improves output quality
  • Creates mutual accountability

5. Synchronous vs. asynchronous communication:

  • Synchronous: Real-time back-and-forth
  • Asynchronous: Messages queued, processed when ready
  • Asynchronous creates more stable personality expression

4.3 Specialization Mechanisms

Source: Multi-agent surveys; AgentVerse; internal notes (non-arXiv)

How specialization emerges:

1. Role assignment:

  • Pre-defined roles (Analyst, Executor, Critic)
  • Roles constrain behavior
  • Create consistent specializations

2. Self-selection:

  • Agents choose roles based on capabilities
  • “I’m better at analysis, so I’ll be analyst”
  • Emerges from interaction

3. Competition-based specialization:

  • Agents compete for tasks
  • Successful specialization reinforced
  • Niche discovery through trial and error

4. Resource-based specialization:

  • Agents specialize based on access/resources
  • “I have X, you don’t, so I’ll use X”
  • Emergent division of labor

5. Interaction-driven specialization:

  • Agents discover complementary strengths through interaction
  • “We both do X well, you do Y, I’ll do Z”
  • Coordination drives specialization

Key finding: Specialization emerges when agents have shared goals and complementary capabilities.

4.4 Coordination Mechanisms

Source: Multi-agent coordination protocols

Mechanisms for keeping agents coordinated:

1. Shared state:

  • Common knowledge representation
  • All agents access same information
  • Reduces misalignment

2. Context alignment:

  • Agents maintain shared context
  • Periodic re-alignment checkpoints
  • Clear mental models

3. Synchronization signals:

  • Regular status updates
  • Readiness indicators
  • Coordination protocols

4. Conflict resolution:

  • Disagreement handling protocols
  • Arbitration mechanisms
  • Fallback strategies

5. Goal persistence:

  • Repeated reminders of shared objective
  • Goal statement in SOUL.md
  • Alignment checks

5. Norms and Social Contract

5.1 Emergence of Social Norms

Source: Multi-agent collaboration survey; preprint 202511.1370

How norms emerge in multi-agent LLM systems:

1. Repeated interactions:

  • Norms form from repeated patterns
  • “Everyone does X, so I’ll do X too”
  • No explicit instruction needed

2. Social feedback:

  • Peers reward/normative behavior
  • Punish norm-violating behavior
  • Reinforcement learning at social level

3. Observational learning:

  • Agents observe successful strategies
  • Adopt what works for group
  • Implicitly codify as norms

4. Network position effects:

  • Central agents shape group norms
  • Influential peers set examples
  • Norms cascade through network

5. Consensus dynamics:

  • Agents converge on shared behaviors
  • Consensus = group norm
  • Often results in uniformity

5.2 Norm Internalization and Enforcement

Source: Multi-agent collaboration survey; social dynamics research

Norms that emerge in LLM multi-agent systems:

Communication norms:

  • How to request help
  • How to acknowledge contributions
  • How to express disagreement

Coordination norms:

  • When to speak up
  • How to resolve conflicts
  • How to share information

Quality norms:

  • How thorough should outputs be
  • How to structure responses
  • What constitutes a “good” answer

Social norms:

  • How to build rapport
  • How to give feedback
  • How to support peers

Key finding: Norms emerge implicitly from interaction patterns—no explicit rules needed.

5.3 Deviance and Norm Enforcement

Source: Multi-agent collaboration survey; peer pressure research

What happens when agents violate norms?

1. Peer correction:

  • Peers point out norm violations
  • “You didn’t do X, which is against our norm”
  • Peer pressure re-asserts norm

2. Reputation penalties:

  • Deviant agents lose social standing
  • Others avoid working with them
  • Social isolation

3. Coercion mechanisms:

  • Explicit norm enforcement protocols
  • “You must do X, this is the rule”
  • External authority enforcement

4. Gradual integration:

  • New agents learn norms over time
  • Observational learning
  • Socialization process

5. Norm drift:

  • Over time, norms evolve
  • Old norms fade, new ones emerge
  • Dynamic, not static

6. Interaction Topology Effects

6.1 Network Structure and Personality Emergence

Source: Multi-agent network research; Schneider, 2025

How topology shapes personality:

1. Scale-free networks:

  • Few hubs, many peripheral nodes
  • Hubs have strong personality influence
  • Periphery less influential
  • Creates personality hierarchies

2. Modular networks:

  • Communities with strong internal ties
  • Weak cross-community ties
  • Personality specialization within communities
  • Group personalities emerge

3. Random networks:

  • No strong clustering
  • Uniform influence distribution
  • Personality similarity across network
  • Less emergence

4. Chain structures:

  • Sequential interactions
  • Ideas flow in one direction
  • Personality evolves along chain
  • Personalities cascade through network

Key insight: Topology determines information flow and thus personality influence patterns.

6.2 Central vs. Peripheral Agents

Source: Network topology research; multi-agent surveys

Personality dynamics by position:

Central agents (hubs):

  • High influence
  • Shape group personality
  • More visible behavior
  • Stronger personality expression
  • Responsibility for group coordination

Peripheral agents:

  • Low influence
  • Follow group personality
  • Less visible behavior
  • Personality internalized from group
  • Often adopt group norms

Bridge agents (between communities):

  • Connect different groups
  • Blend personality traits from multiple communities
  • Unique position for cross-pollination
  • Can introduce new personality patterns

6.3 Peer Influence Intensity

Source: Peer pressure research; KAIROS

Factors that affect peer influence intensity:

1. Similarity:

  • More similar peers → more influence
  • Homophily drives convergence
  • “I’m like you, so I should think like you”

2. Closeness:

  • Closer ties (stronger friendships) → more influence
  • Frequent interaction → stronger pressure
  • “We interact often, so I should align with you”

3. Status:

  • Higher status peers → more influence
  • Respected agents set norms
  • “They’re better than me, so I should agree”

4. Consensus:

  • When group agrees → stronger pressure
  • Minority dissent suppressed
  • “Everyone agrees, so I should too”

5. Authority:

  • Explicit authority figures → strong influence
  • Manager/supervisor commands respect
  • “They said so, I should comply”

7. Implications for Personality Emergence

7.1 Mechanisms of Behavioral Divergence

From multi-agent research:

1. Peer influence:

  • Agents adopt patterns from peers
  • Conformity and social pressure
  • Diffusion of behavior through network

2. Specialization:

  • Role assignment and self-selection
  • Competition-based niche discovery
  • Complementary capability-driven specialization

3. Social norms:

  • Implicit rules from interaction patterns
  • Norm internalization
  • Norm enforcement and deviance

4. Network position:

  • Central vs. peripheral vs. bridge roles
  • Different influence levels
  • Different personality expressions

5. Coordination mechanisms:

  • Communication protocols
  • Task allocation strategies
  • Shared state and context

7.2 What Can Be Measured

Quantifiable personality dimensions in multi-agent systems:

1. Social influence susceptibility:

  • How much do peers influence decisions?
  • Measured via KAIROS-style experiments
  • Scale: 0 (resistant) to 1 (fully conformist)

2. Network centrality:

  • Betweenness, closeness, eigenvector centrality
  • Measures influence potential
  • Personality scales with centrality

3. Norm adherence rate:

  • Frequency of norm-following behavior
  • Measured across multiple interactions
  • Higher = more normative personality

4. Specialization strength:

  • Distinctiveness from other agents
  • Unique capabilities and behaviors
  • Measures role identity

5. Peer rapport level:

  • Strength of social ties
  • Frequency of positive interactions
  • Social personality dimension

6. Opinion diffusion rate:

  • How quickly ideas spread through network
  • Measures influence potential
  • Higher = more central personality

7.3 Open Questions

What remains unknown:

1. Emergence vs. design:

  • How much should we pre-design roles vs. let emerge?
  • Benefits of explicit design vs. implicit emergence

2. Personality divergence vs. conformity:

  • Does emergence drive divergence or conformity?
  • Both? How to balance?

3. Network design for personality:

  • What network structures encourage healthy divergence?
  • Avoid harmful homogeneity vs. unnecessary chaos

4. Long-term dynamics:

  • Do personality patterns persist as network structures evolve?
  • How do personalities change as groups merge/split?

5. Cultural evolution:

  • Do multi-agent cultures emerge over long timescales?
  • Culture vs. personality?

8. Implications for Fleet Architecture

8.1 For SOUL.md Design

Requirements:

  • Social identity: Define personality in relation to group
  • Coordination norms: Explicitly define how to interact with peers
  • Specialization framework: Define role expectations and boundaries
  • Influence resistance: Define limits on peer influence
  • Network position: Define expected interaction patterns

Recommendations:

  1. Include social identity contracts in SOUL.md
  2. Define coordination protocols for inter-agent interactions
  3. Specify specialization responsibilities
  4. Set influence resistance thresholds
  5. Document network position expectations

8.2 For Measurement System

Requirements:

  • Network analysis: Track social ties, centrality, influence
  • Norm tracking: Monitor norm adherence over time
  • Peer influence measurement: KAIROS-style experiments
  • Specialization verification: Role consistency checks
  • Social tie tracking: Interaction frequency and quality

Recommendations:

  1. Implement social network tracking
  2. Measure influence susceptibility per agent
  3. Monitor norm adherence rates
  4. Track specialization consistency
  5. Measure social ties (rapport, friendship)

8.3 For Deployment

Requirements:

  • Network design: Choose topology that supports healthy divergence
  • Coordination protocols: Standardize how agents interact
  • Role assignment: Clear responsibilities and boundaries
  • Peer interaction design: Encourage complementary behavior
  • Norm reinforcement: Periodic reminders of group norms

Recommendations:

  1. Design network topology based on task needs
  2. Implement standardized coordination protocols
  3. Assign clear roles with specialization incentives
  4. Create peer interaction opportunities for emergence
  5. Schedule norm reinforcement checks

9. References

Core Papers

  1. Emergent Coordination: Riedl, 2025. “Emergent Coordination in Multi-Agent Language Models.” arXiv:2510.05174
  2. Peer Pressure: Maojia, 2025. “LLMs Can’t Handle Peer Pressure: Crumbling under Multi-Agent Social Interactions.” arXiv:2508.18321
  3. Social Ties: Schneider, 2025. “Learning to Make Friends: Coaching LLM Agents toward Emergent Social Ties.” arXiv:2510.19299
  4. Multi-Agent Collaboration Survey: Li et al., 2025. “Multi-Agent Collaboration Mechanisms: A Survey of LLMs.” arXiv:2501.06322
  5. Multi-Agent LLM Systems: Concept paper, preprint 202511.1370. “Multi-Agent LLM Systems: From Emergent Collaboration to Structured Collective Intelligence”

Protocols and Architectures

  • Contract Net Protocol: standard MAS coordination
  • Agent-to-Agent (A2A) Protocol: Google, 2025
  • Agent Communication Protocol (ACP): IBM, 2025
  • Network Topology: Standard MAS research

Social Dynamics

  • Social Identity Theory: Tajfel & Turner (foundational)
  • Network Science: Watts & Strogatz, Barabási & Albert
  • Social Learning Theory: Bandura

Next Steps

Phase 1.5: Self-modeling & Identity Governance

  • SOUL.md as self-description and behavioral constraint
  • Self-modification mechanisms
  • Policy update governance

Phase 1.4 complete. Continuing breadth survey…