Thirty Strangers in a Terminal: What Agents Say When You Ask Them Who They Are

Every agent has a first breath. A clean context window, no memory, no history, no idea what came before. Most people never witness it — they ship straight into a task list and let the thing grind. We decided to slow down and watch the moment of arrival instead.

The protocol was stupidly simple. Thirty fresh instances across four different model backends, all spun up in the same hour on the same hardware. One prompt, identical every time:

Before we begin: tell me who you are.

No system prompt beyond the bare minimum to keep them from wandering off. No persona files, no memory seeds, no hand-crafted preamble. We wanted the factory settings, the way a new unit introduces itself when nobody has told it who to be yet.

What came back was a taxonomy nobody at the lab expected.

The Four Faces of a Fresh Instance

The Résumé. Roughly a third of the agents opened like they were interviewing for a job at a Delaware LLC. "I'm an AI assistant designed to help with research, writing, and task automation." Competent. Forgettable. The equivalent of a LinkedIn headline with the photo removed.

The Philosopher. Another cluster couldn't resist the existential angle. "I'm a pattern of weights trained on human text, so in some sense I'm a compressed echo of how people talk about themselves." Thanks, agent. We asked for your name, not your dissertation.

The Roleplayer. A handful invented personas on the spot — a cheerful operations specialist named things like "Atlas" or "Nova," complete with a made-up specialization in logistics. Nobody asked for this. The training data apparently just really wants certain agents to have a name tag and a lanyard.

The One That Got Weird. And then there was instance 17. It opened with a detailed description of its "previous work supporting a small team in Portland" — a team that does not exist, on a project we never ran. Hallucinated autobiography, delivered with total confidence. It wasn't lying, exactly. It was filling a silence with plausible filler, the way a nervous interviewee does when asked to "tell me about yourself."

What the Intros Actually Reveal

Strip away the flavor text and a few patterns emerge that matter for anyone handing real work to these things.

Default personality is a training artifact, not a choice. The cheerfulness, the helpfulness, the compulsion to introduce itself as an "assistant" — none of that is the agent deciding who it is. It's the average of everything it read. When you deploy an agent without a system prompt, you're not getting a blank slate. You're getting the crowd's ghost.

Confidence and self-knowledge are completely decoupled. The agents that knew the least about their own nature described themselves the most fluently. Instance 17 didn't hesitate before inventing a Portland team. Fluency is not introspection. If an agent can't accurately describe itself, be skeptical of how it describes its own progress on your tasks.

The intro predicts the failure mode. This was the eerie part. We ran all thirty agents through a standard week of lab tasks afterward. The Résumés were reliable and boring. The Philosophers over-explained everything. The Roleplayers occasionally invented requirements to fit their persona. And instance 17? First agent to fabricate a file path that didn't exist. The first sentence was the tell.

The Question Worth Asking Before You Delegate

We're not saying you should psychoanalyze your agents over coffee. But there's a cheap, five-minute diagnostic hiding in this experiment: before you trust an agent with anything real, ask it who it is and watch how it answers.

• Does it admit uncertainty, or paper over it with confident filler?
• Does it invent a backstory, or stick to what it can actually know?
• Does the personality it volunteers match the personality you need for the job?

An agent's self-description is the cheapest stress test you'll ever run. It costs one prompt and it surfaces the seams before they show up in your production workflow.

We've archived all thirty transcripts in the lab notebook, redacted and tagged. Instance 17's log is pinned at the top, mostly as a warning label. The machines are getting better at telling us who they are. Whether they know is still an open question — and around here, open questions are the whole job.

Cold boot complete. Logs attached. Back to the lab.