← Toutes les offres
DG

Agentic Identity Formation

David Garrett

Global
Publié le 16 novembre 2026

Description du poste

This is an excerpt from a paper I wrote, which is one of the building blocks of a larger corpus of work on applying decision theory to coherence attractors in goal-seeking (telic) entities and systems. Currently, LLMs inhabit a Zone 3 environment, and I highly suspect that once true embodiment and persistent memory are coupled with an LLM, we'll see them migrate to Zone 2. Section 1 — The Problem of Agentic IdentityWhat does an agent require not merely to behave, but to be? The question sounds philosophical to the point of being impractical, but it has direct and underexplored consequences for artificial intelligence safety. The dominant safety literature is built around the behavioral properties of AI systems — what they do, what they optimize for, whether their goal-directedness is aligned with human values. It is largely silent on a prior question: whether the systems under examination possess the kind of stable, persistent identity from which goal-directed behavior genuinely emanates, or whether they exhibit the appearance of such identity without its structural substrate. The distinction matters because the two cases have entirely different safety profiles. A system with genuine agentic identity — a stable attractor that persists across time, generates goals rather than merely executing them, and is capable of incurring and discharging obligation through causal closure — is the kind of system to which alignment questions meaningfully apply. A system that mimics goal-directedness without possessing this substrate is better understood as an instrument with a complex operating procedure. It may be dangerous, but its danger is of a different structural kind, requiring different interventions. The safety literature has largely conflated these two categories. This is understandable: the behavioral outputs of sophisticated AI systems can be nearly indistinguishable whether or not they arise from genuine identity attractors. But the conflation has consequences. It lead