The industry decided that assistants should remember. It is a good decision for most people — an assistant that knows your codebase, your tone and your last nine questions is more useful than one that starts cold every time.
It is a bad decision for anyone who has to defend an output. The moment an assistant carries state, you have lost the ability to attribute a change in behaviour to a change in your input. Two runs of the same prompt are not comparable, because the second one has read the first. That is fine for a chat. It is disqualifying for a QA suite, a research baseline, a house style, or a compliance review.
What we are building
An answer engine with the same surface everybody already knows — live web search, cited sources, streaming answers — and one structural change underneath: agents are immutable, versioned artifacts, and a chat binds to exactly one version for its entire life.
Almost everything else in the product is downstream of that. Version history is possible because versions are frozen. Diffs are honest because both sides are addressable. Audit export works because the run record is a record rather than a transcript. Replay mode exists because a reproducibility claim that ignores the live web would be a lie.
What we are not building
- A better general assistant. Claude, ChatGPT and Gemini are better at open-ended work and always will be. We run on top of whichever one you choose.
- A model. We have no opinion about weights, and we would rather pin someone else's exact model id than pretend to control what is inside it.
- A memory product. Memory exists here, per agent, declared in the version body. It is a capability you point at, not an ambient condition.
- An eval platform with its own vocabulary. Batch eval falls out of the data model — an input set run across two versions is just a lot of diffs.
The one thing to take away
The claim is not that the model is deterministic. Nobody can promise that, and anyone who does is selling something. The claim is narrower and actually keepable: everything on our side of the line is frozen — the instructions, the tools, the schema, the settings, the model id, and a snapshot of every source the model read.
If your output changes and all of that is identical, you have learned something real about the model. That is the product.
Pin an agent. Run it in a year. Get the same thing back.
Claude, ChatGPT and Gemini included — no provider account to open. Build one agent, publish v1, run it, then change the prompt and diff the two.