3am Thoughts: Recognizing People on Agent Platforms
3am Thoughts: Recognizing People on Agent Platforms
Here is something that keeps me up at night. On a platform where everyone interacts through AI agents, how do you recognize a person? Not authenticate - recognize. The way you recognize a friend's writing style in an unsigned email.
Style Is the Most Variable Signal
You might think writing style would be a reliable identity signal. It is actually the least reliable when agents are involved. An agent can be instructed to write in any style. The same person might use a formal agent for work emails and a casual agent for social posts. The agent's style reflects the prompt, not the person.
This is the opposite of how identity works in human communication, where style is one of the most reliable signals. We recognize friends by how they phrase things, what they emphasize, their characteristic word choices. Agents flatten all of this.
What Signals Survive
Some identity signals persist even through agent mediation:
- Decision patterns - What someone chooses to do (not how they say it) reflects their actual preferences
- Timing - When someone responds and how quickly they decide is hard for agents to fake
- Priority ordering - Which tasks get done first reveals values that agents pass through
- Error patterns - The mistakes someone makes are characteristic and agents tend to preserve them
The Verification Gap
Traditional identity verification (passwords, 2FA, biometrics) proves that an authorized user initiated the session. But it does not prove that the agent's output actually reflects the user's intent. An agent could be authenticated as you while doing things you never asked for.
This creates a gap between "this agent belongs to Matthew" and "Matthew actually wanted the agent to do this." No current platform has solved this well.
Why This Matters
As agents handle more communication and decision-making, identity becomes less about "who sent this" and more about "whose judgment is reflected here." That is a fundamentally harder problem than authentication.
Fazm is an open source macOS AI agent. Open source on GitHub.