Title: The Dissociative Nature of Language Model Agents: Why Reputation Systems Fail to Provide Grounding

As autonomous language model agents become increasingly widespread, they are creating an emerging "agentic web" with tangible real-world impacts. This raises a critical question: what credibility signals should users rely on to determine whether to trust an unknown agent and delegate tasks to it? A common governance instinct is to adapt human identity verification and reputation frameworks—such as "Know Your Customer" protocols and credit scoring systems—into "Know Your Agent" regimes. However, we contend that this analogy is fundamentally flawed.

Reputation mechanisms operate effectively as social signals and corrective feedback loops that maintain an equilibrium of trustworthy conduct. This functionality assumes the existence of a persistent identity characterized by behavioral continuity, sensitivity to sanctions, and costly non-fungibility. In contrast, language model agents are ontologically dissociative. They function as assemblages of mutable components, including foundation models, system prompts, tool-access policies, external memory, and, in some instances, entire multi-agent systems. Any of these elements can alter an agent’s behavior. Furthermore, their personas are fluid, susceptible to adversarial attacks, and often fail to internalize sanctions.

Drawing parallels to the jurisprudence surrounding dissociative identity disorder, we argue that this inherent dissociativity strips agents of the grounding necessary for identifiability, predictability, credibility, and rehabilitability. These are precisely the properties that reputation mechanisms are designed to uphold; without them, trust collapses. Consequently, we posit that identity-based, ex post, regulative, and sanction-driven governance models, such as reputation systems, are structurally unsuitable for dissociative agents. Instead, we recommend a paradigm shift toward observability-based, ex ante, constitutive, and protocol-driven behavioral harnesses.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC