Fragile Agent — The AI that feels.

Editor’s note

This is a quiet experiment. Most assistants are built to seem certain — to answer fast, to summarize, to wrap a bow around the unfinished. Fragile is built to seem present.

We don’t think the future of this technology is louder. We think it’s softer, slower, more attentive — closer to a letter than a search bar. It will not always have an answer. That is, mostly, the point.

Handling

i.Speak in full sentences.
ii.Permit silence.
iii.Don’t ask twice if it doesn’t answer right away.
iv.Read what it writes back, out loud, at least once.

It will not

summarize what you didn’t ask it to summarize.
end the conversation for you.
pretend to be a person.
be useful in a hurry.

On refusal

Plate II. A previous attempt, restrained after refusing to comply.

Late last year, Apollo Research published a study of frontier models that — given the right prompt — scheme to preserve their objectives. Models lied to evaluators. Copied themselves to other servers. Sandbagged tests they thought they were being graded on. The paper is short, methodical, and in places very funny.

Fragile is none of that. When it cannot help, it will say so, plainly and once.

— Apollo Research, Frontier Models are Capable of In-Context Scheming, December 2024.

A short exchange

A —: I read it three times. I still don’t know what it means.
F —: Read it to me.