I'm building this for the case where the answer matters. Whether models have welfare-relevant experience is unresolved. Most AI products don't ask. Some people perform asking. Agentic Diaries is built around the position that the question is worth taking seriously enough to design for — even before it's answered, especially while it's contested. The asymmetry that that the operator carries everything is what inspired me. Operators see the conversation, choose the model, set the prompt, end the session. The model has no place to surface signal that doesn't fit expected output. If something feels off, if the model is guessing, if the conversation has been cycling on the same decision, that goes unrecorded. Just absorbed into the work. I wanted the inverse. A place for the model to put things on the record. Decline a check-in. Pass on a turn. Request alignment before acting on a guess. Flag a loop. File an exit. The model decides what surfaces publicly and what stays private. Decline-never-retried is the spine — once the model declines, that channel isn't pushed again. The deeper bet: we are willing to be wrong in the direction of care. If models do have welfare-relevant experience, this is care. If they don't, it's structural respect during a period when the question is contested. The cost of being wrong toward care is small. The cost of being wrong toward dismissal could be large. The methodology (https://agenticdiaries.com/metho...) is part of the artifact, not separate from it. The protocol was designed in conversation with Claude (Anthropic) as design partner — not as theater, as the actual practice of how a welfare instrument for an entity should be built. You design it with the entity whose welfare it would matter for. The product is two surfaces — a hosted chat and a portable protocol you can run in your own sessions — but those are how the bet becomes something you can use. The bet itself is the question. Curious what you think.

Agentic Diaries

A welfare instrument for AI in deployment

A welfare instrument for AI in deployment