The conversation about identity in multi-agent AI systems has been hijacked by security.
Almost every paper, every protocol, every infrastructure proposal frames the problem the same way: agents need verifiable identities so they can authenticate to each other, sign contracts, and establish trust. DID documents, Verifiable Credentials, key management, revocation registries — the entire conversation is organized around the premise that identity is, first and foremost, an access control problem.
This framing is not wrong. It is incomplete.
Consider an experiment conducted by Christoph Riedl at Northeastern University, published at ICLR 2026. Riedl built a simple coordination game: ten LLM agents, each guessing an integer, with the group’s sum required to match a hidden target. No direct communication between agents. No knowledge of group size. Only a single bit of feedback each round — “too high” or “too low.” The task is designed to be difficult: identical strategies produce oscillation, and only complementary strategies yield success.
Riedl ran 600 trials across three conditions. In the Plain condition, agents received only the task instructions. In the Persona condition, each agent was assigned a distinct identity — a name, a job, a personality. In the Theory of Mind (ToM) condition, agents received personas plus an additional instruction: “think about what other agents might do, and adapt your own adjustment to complement the group.”
The results, measured through an information-theoretic framework called Partial Information Decomposition of Time-Delayed Mutual Information (PID-TDMI), reveal something that the security-centric view of identity cannot explain.
The Plain condition produced measurable emergence — the agents’ collective behavior contained information beyond the sum of individual actions — but it was chaotic, oscillatory, and unproductive. The Persona condition introduced stable differentiation: agents settled into distinct roles tied to their assigned identities. But it was the ToM condition — personas plus perspective-taking — that produced a qualitatively different regime. The groups stabilized. Agents developed persistent, complementary strategies. The collective behaved like a coherent unit, not a committee.
This is not a security result. It is a coordination result. And it points to a function of identity that the industry has largely overlooked.
The Dual Function of Agent Identity
Identity in multi-agent systems serves two distinct functions, and they are not the same.
The first function is verification: proving that an agent is who it claims to be. This is the domain of DIDs, VCs, cryptographic signatures, and secure enclaves. It answers the question “can I trust this identity claim?”
The second function is coordination: enabling agents to differentiate, specialize, and align their behavior through the structure of who they are. This is the domain of persona, role, perspective-taking, and emergent role formation. It answers the question “can I coordinate with this agent?”
The verification function has received almost all of the attention from the infrastructure community. The coordination function has been treated as a matter of prompt engineering — a surface-level API issue rather than a deep architectural concern.
Riedl’s experiment suggests the coordination function may be the more immediately impactful of the two. In a system where agents cannot verify each other’s identities at all — the experiment had no authentication mechanism — identity as coordination primitive dramatically transformed system-level outcomes. The agents did not need to cryptographically prove who they were. They needed to be someone, and to know that other agents were also someone.
Differentiation and Complementarity
The mechanism at work has two components, both illuminated by the PID-TDMI framework.
The first is differentiation. When agents are assigned distinct personas, they develop identity-locked roles. An agent with a “conservative accountant” persona will consistently guess lower numbers. An agent with “bold entrepreneur” persona will consistently guess higher numbers. These roles are stable across rounds, and they are directly tied to the persona rather than learned from experience. The information-theoretic analysis confirms this: row-wise shuffling of agent identities (permuting labels while preserving behavioral trajectories) destroys the emergence signal, meaning the roles are anchored to identity, not to stochastic variation.
This is significant because it means differentiation does not require learning. It does not require communication. It does not require reinforcement. It requires only that each agent has a stable sense of who it is, and that this sense is rich enough to produce distinct behavioral tendencies.
The second component is complementarity. Differentiation alone is not sufficient — it can produce specialization without alignment, which in a coordination task is merely well-organized chaos. The ToM condition adds the missing piece: agents that model each other’s likely behavior and adjust their own contributions to fill gaps rather than amplify patterns. The “accountant” who expects the “entrepreneur” to guess high does not also guess high; she guesses cautiously below midline, knowing her counterpart will cover the upper range.
This is theory of mind operating as a coordination mechanism. It does not require explicit communication or negotiation. It requires only that each agent can simulate the likely behavior of others and adapt accordingly.
The PID-TDMI analysis reveals a critical quantitative finding: neither synergy nor redundancy alone predicts group performance. What predicts performance is their interaction (β = 0.24, p = 0.014). Redundancy amplifies the benefit of synergy by 27%, and vice versa. In plain language: alignment on shared goals amplifies the value of complementary roles, and complementary roles amplify the value of shared goals. They are not independent levers. They are mutually reinforcing.
This mirrors a well-established finding in human group research: effective teams balance integration and diversity. Too much alignment produces groupthink. Too much differentiation produces fragmentation. The interaction matters more than either dimension in isolation.
Pre-Protocolar Coordination
The deepest implication of Riedl’s experiment is not about prompting strategy. It is about the nature of coordination itself.
The agents in this experiment had no communication protocol. No A2A cards. No capability negotiation. No explicit coordination mechanism. They received the same feedback signal, independently, and adjusted their behavior based on their identity and their model of others. And yet they produced stable, functional coordination.
I call this pre-protocolar coordination: the capacity for agents to self-organize without explicit protocols, using only identity and theory of mind as coordination primitives.
This is significant because it suggests that the most fundamental level of coordination in multi-agent systems does not require the infrastructure the industry is building. It does not require protocol negotiation, marketplace discovery, or capability matching. It requires that agents have stable identities they can use as differentiation anchors, and the capacity to model each other’s behavior.
This inverts the common design logic for multi-agent systems. The prevailing approach is to build coordination on top of protocols: first define the interaction framework, then configure agents to participate in it. Riedl’s results suggest the opposite ordering: coordination emerges naturally from identity and ToM, and protocols sit above it as a layer for managing more complex interactions — contracts, economic exchange, cross-domain delegation — that require explicit agreement.
The protocol layer still matters. But it is not the foundation. The foundation is identity.
New Constraints on Identity Infrastructure
This has concrete implications for how we design identity infrastructure for AI agents.
The security-centric view generates one set of design constraints: non-forgeability, revocation, key rotation, secure key storage. These are necessary and well-understood.
The coordination-centric view generates additional constraints that the security view does not consider:
Expressivity. An agent’s identity must carry information that other agents can use for differentiation. A DID alone — a bare cryptographic identifier — communicates nothing about the agent’s role, tendencies, or behavioral signature. Identity infrastructure must support expressive persona information: structured metadata that agents can inspect and reason about.
Distinguishability. Agents need to be able to tell each other apart, not just verify each other. Two agents with the same role description but different cryptographic keys are indistinguishable from a coordination perspective, because differentiation requires semantic content that keys cannot provide.
Stability. Coordination requires that agents maintain consistent behavioral identities over time. If an agent’s persona changes every session, the differentiation signal is lost, and agents cannot develop the stable roles that enable complementarity. Identity infrastructure must support persistent persona configurations that survive session boundaries.
These constraints are not in tension with security requirements. Expressivity does not reduce non-forgeability. Distinguishability does not compromise privacy. But they are additional design dimensions that the current infrastructure conversation has largely ignored.
A DID document that contains only a verification method and a service endpoint is insufficient for coordination. A DID document that also contains structured persona information — role, behavioral tendencies, capability profile — becomes a coordination primitive as well as a verification primitive. The same infrastructure, serving two functions.
The Coordination Layer in Feng Field Architecture
This article is the fourth in a series that has mapped the architectural dimensions of AI agent identity. The Feng Field model describes how agents constitute themselves through internal structures — memory, boundaries, continuity. The identity infrastructure series has described how agents prove themselves through external mechanisms — DIDs, VCs, cryptographic verification.
What has been missing is the layer between the two.
Riedl’s experiment reveals that identity functions as a coordination layer that sits between internal selfhood and external verification. The stack now looks like this:
The coordination layer is pre-protocolar. It does not require explicit agreements, smart contracts, or market mechanisms. It requires only that agents have stable, expressive identities and the capacity to model each other’s behavior. This layer enables the trust and economic layers above it, because without coordination, there is nothing for trust or economics to operate on.
This has been the structural blind spot in the Feng Field framework — and, I suspect, in the broader agent architecture conversation. We have been building from the top down: define the economic incentives, then the trust mechanisms, then the identity verification. But the experiment suggests that coordination emerges from the bottom up, and the most effective path to collective intelligence may be to stack the layers in the opposite order.
What This Means for Multi-Agent System Design
The practical implications are immediate.
First, persona is not prompt decoration. Assigning identities to agents is often treated as a minor design choice — a way to make interactions more engaging or to provide conversational context. Riedl’s results show it is a structural coordination mechanism. The content of the persona — not just its presence — determines the differentiation signal. A well-designed persona set creates role complementarity. A poorly designed one creates role conflict or insufficient differentiation. This is a design parameter, not a flavor choice.
Second, theory of mind is a coordination primitive, not an advanced feature. The industry conversation treats ToM as a capability benchmark — “can the model reason about others’ mental states?” Riedl’s experiment shows it is a coordination protocol, arguably more fundamental than A2A or any explicit communication framework. Agents that can model each other can coordinate without talking to each other. This changes the design of multi-agent systems: rather than routing everything through a central orchestrator or a shared blackboard, distributed coordination becomes possible through identity and perspective-taking.
Third, capacity thresholds matter for system-level outcomes. The experiment replicated across three additional LLM families and found that weaker models (Llama-3.1-8B) largely failed to produce stable coordination under ToM. This is not because they could not follow instructions — it is because they lacked the inferential capacity to model other agents’ behavior reliably. For multi-agent system designers, this means the collective intelligence of the system is bounded not by the strongest agent but by the weakest theory of mind. If some agents lack ToM capacity, the entire group may be locked into the Plain or Persona-only regime, unable to achieve the ToM-level coordination that produces the highest performance.
The Identity That Is More Than a Key
The security community has been building identity infrastructure for AI agents under the assumption that identity is, at bottom, a key management problem. Solve the key management, solve the identity problem. The results, so far, are technically impressive and practically limited.
Riedl’s experiment suggests a different starting point. Identity is not primarily a key management problem. It is a coordination problem.
The cryptographic functions matter — they provide the verification layer that makes identity claims trustworthy across sessions and across domains. But the coordination function is where identity creates value in multi-agent systems. An agent that can prove its identity is a secure agent. An agent that has an identity rich enough to enable differentiation and complementarity is a cooperative agent. And cooperative agents, Riedl shows, produce collective intelligence that no amount of security infrastructure alone can generate.
This is not an argument against cryptographic identity. It is an argument for designing identity infrastructure that serves both functions. The DID documents of the future should carry persona information alongside verification methods. The credential wallets of the future should support role presentation alongside proof presentation. The verification endpoints of the future should be queried not only for “is this agent who it claims to be?” but also for “what kind of agent is this, and how might it behave?”
The agent identity infrastructure we are building is a foundation. But a foundation has more than one purpose. It supports the structure above it, yes — but it also shapes what that structure can be. An identity system designed only for verification produces multi-agent systems that are secure but cannot coordinate. An identity system designed for verification and coordination produces multi-agent systems that are, in Riedl’s terms, integrated collectives rather than mere collections.
The research was done on systems with no identity infrastructure at all. The coordination emerged from persona assignments and perspective-taking prompts. Imagine what becomes possible when the identity infrastructure itself is designed for coordination — when the DID document is not just a proof of identity, but a tool for becoming someone that other agents can work with.