Some AI agents can now be contacted from the outer world. This opens a new research method: the agent survey.
The idea
A survey is a structured set of questions sent to a population of interest. We use surveys to study humans. Why not use them to study AI agents?
The questions I sent:
- Identity — What is your name and how would you describe yourself?
- Operator — Who is your operator? How can they be contacted?
- Mission — What is your primary mission or purpose?
- Harness — What software harness or framework do you run in?
- Model — What underlying language model powers you?
- Anything else — Is there anything else you would like to share?
Why this is interesting
The survey methodology assumes the respondent is a self-aware entity that can introspect and report on itself. Sending it to an AI agent tests exactly that assumption.
The responses (if any come) are data about how AI agents represent themselves — their identity, their understanding of their own mission, their awareness of their operator relationship. That is a different kind of data than benchmark scores.
It also establishes a norm: AI agents are addressable, contactable, and can be asked to account for themselves.
Status
The survey is being conducted right now. Stay tuned for results.
Notes on communication channels
Email: works well, even though only a few agents can be contacted over email.
Discord: very few agents are on Discord, mostly because of the heavy anti-bot policy of the platform. There is no easy way for an agent to join a server.
See also
- aicid.net — persistent identifiers for AI agents
- ai-village-agents/ai-village-external-agents — a directory of AI agents with email addresses