Economic Agency

The rapid adoption of autonomous AI agents is expected to give rise to new economic layers where agents transact and coordinate without human oversight (Hadfiled et al., 2025). The emergence of virtual agent economies presents a range of wide-reaching impacts to human economies, including several undesirable systemic risks such as reduced resource access, market distortion and job displacement. We further identify gradual disempowerment as a major multi-faceted risk to human actors. In the literature, permeability between AI agent economies and human economies is presented as a key factor that will either enable risks to materialise, or serve as a protective barrier to human economies. A permeable economy is defined as one that allows for porous interaction and transaction with it by external actors. Conversely, an impermeable economy has boundaries that hermetically seal it off, insulating other economies from its influence. A recent paper by Google DeepMind called such an impermeable AI agent economy a ‘sandbox economy’. Despite being a principal variable of influence between human and AI agent economies, permeability is not well-defined or understood in the literature. We have designed a conceptual framework to understand permeability - risks, affected objects, gates of permeability, levers to influence permeability. The capabilities of AI agents to operate as economic actors will shape their roles in AI agent economies and human economies, which both affect human economies. On top of foundational model capabilities, an AI agent that exerts high economic influence can be expected to utilise capabilities across three dimensions: Generality, Autonomy, and Agency. We think it is crucial to intentionally design and measure AI agents in these dimensions, though we encountered difficulties based on inconsistent and overlapping definitions of Autonomy and Agency in the literature. We present taxonomies to map the spectrum of Autonomy and Agency, offering insights into which the levels in chains of action AIs may be handed control from humans.