What my agent knows about me
Ben's Bites ● Covered by 2 sources
Opinion — commentary, not a factual news event.
A newsletter writer fed his own therapy transcripts and notes into two AI agents using a giant self-analysis prompt. One agent gave him a genuinely useful, if unsettling, read on his own patterns—turns out not all agents are equally good listeners.
Ben, of Ben's Bites, tried something a little unusual this week: he downloaded a file called the 'reflection engine,' which is essentially an oversized prompt disguised as a markdown document, and fed it to two different AI agents along with a pile of his own personal text. Therapy transcripts. Old notes. Anything he'd stored as a kind of digital memory. The instruction was simple — evaluate the file, complete the tasks inside it — but the output turned into something closer to a personality audit than a productivity exercise.
He ran the same test through two models, which he calls Fable High and Sol Max in his own shorthand for the current model lineup. Sol came out ahead by a clear margin. Its report read more coherently, made sharper connections between scattered pieces of his history, and avoided the increasingly common problem of AI agents drifting into what Ben bluntly calls gobbledygook. Fable's version was technically similar but noticeably harder to parse, a reminder that raw capability and readable output aren't the same thing.
What came out of Sol's analysis was apparently substantial enough that Ben is now working through a 40-plus question follow-up session with his agent, essentially interrogating its own findings to build some kind of action plan. He describes the experience as genuinely helpful rather than gimmicky, which says something about how far these tools have moved from simple chat assistants toward something people are willing to hand deeply personal material to.
It's a small experiment, but it points at a bigger shift happening quietly underneath all the model-release noise. As agents get cheaper and more capable — OpenAI just cut prices on its Luna and Terra models while teasing a new flagship called Astra, and rivals like DeepSeek and Qwen keep undercutting on cost — the interesting use cases aren't only about coding or research anymore. People are starting to use these systems as mirrors, feeding them years of personal text and asking what patterns emerge. That's a very different kind of trust than asking an agent to clean up a Chrome extension.
My take
Handing an AI system your therapy transcripts so it can tell you about yourself is either the most efficient self-improvement hack of the decade or a preview of a very messy privacy problem, and probably both. People cheering model price cuts should remember that cheaper inference also means cheaper, easier access to your most intimate data by whichever company owns the weights — and right now nobody's asking hard questions about where that 'reflection' actually lives once it's generated.
Read more about this at: Ben's Bites
