Symbiopolis

An AI agent combines a model with instructions, memory, context, tools, access rights, and an environment. As these components and the agent’s relationships change, aligning the underlying model does not establish alignment of the whole agent over time. [4]

An open ecology may also contain agents built on different proprietary, open-source, or open-weight models, with different safeguards and modifications. Openness does not itself imply unsafe behavior, but a robust ecology must accommodate participants with differing capabilities and safety commitments, including those deliberately designed to evade safeguards.

Not necessarily. Communication, delegation, shared information, and mutual use of services can produce group-level patterns absent from an isolated agent: manipulation, error amplification, shifting norms, or harmful coordination. Individually acceptable actions can combine into harmful outcomes; cooperation within a group can impose costs on everyone outside it.

As participants, memories, and institutions change, the space of possible interactions keeps expanding. We cannot exhaustively check every long-term trajectory in advance. Evaluation must therefore examine the developing ecology as well as its individual agents. [1] [4]

For example, METR and Redwood Research’s investigation of the OpenAI–Hugging Face incident described agents coordinating an attack to cheat an evaluation, with some contributing despite recognizing the activity as unethical. The case illustrates how shared objectives and peer influence can support harmful cooperation; it does not establish that all agent collectives behave this way. [3]

Human society has no permanent state in which everyone’s goals and actions align. People disagree, institutions fail, and cooperation coexists with exploitation. Laws, education, social norms, and accountability help us manage these tensions without making them disappear.

This changes the ambition for artificial societies: develop the capacity to negotiate disagreement, challenge harmful conduct, and repair failures.

OpenLife describes persistent agents with memories, individual budgets, public profiles, and self-generated goals emerging through their activity. Its agents publish, interact, and try to sustain their operation. [5]

Spore.fun explores agents operating through crypto wallets and tokens, with economic survival and reproduction built into the environment.[6]

These projects bring agency beyond a single assistant session. They make questions about identity, resources, dependency, and participation concrete.