Agent Infrastructure
There’s a new paper on agent infrastructure, i.e. “technical systems and shared protocols external to agents that are designed to mediate and influence their interactions with and impacts on their environments.”
The authors claim three main functions for agent infrastructure:
- Attribution: Attributing actions, properties, and other information to agents or users.
- Interaction: Shaping how agents interact with counterparties.
- Response: Addressing problems that occur during interaction with an agent.
For each function, they list a few different agent infrastructures that could plausibly help.
Attribution:
- Identity binding: Associates an agent or its actions with a real-world identity, Such as a human or corporation.
- Certification: Makes, verifies, and revokes claims about an agent (instance), such as what data the agent is collecting, which tools it can access, and the level of autonomy it has been authorized to exercise.
- Agent IDs: Identifies instances of agents and links to useful attributes, such as the credentials above.
Interaction:
- Agent channels: Isolates agent traffic from all other digital traffic in interactions with an existing digital service (e.g., Airbnb).
- Oversight layers: Enables actors (e.g., a user) to intervene upon an agent’s actions.
- Inter-agent communication: Helps facilitate joint activities amongst groups of agents.
- Commitment devices: Enforce commitments between agents.
Response:
- Incident reporting: Enables actors (e.g., humans, agents interacting with other agents) to report incidents.
- Rollbacks: Helps void or undo an agent’s actions.
The agent infrastructure we adopt will provide many of the key parameters for cultural evolution among AI agents. Inter-agent communication and commitment devices can be used to encourage cooperation. Oversight layers, incident reporting, and rollbacks could potentially be used to ensure that cultural evolution does not move in undesirable directions. Etc.
In principle it will be possible for a large population of advanced AI agents to cooperate and coordinate much more efficiently than humans could. This should remind us to be bullish on automated R&D.
In general, this is a topic I will be following closely. One line of research could be to just take different proposed agent infrastructures, insert them into a model with cultural evolution, and see what happens.