Avagenc

Every assistant I used could tell me what was in my calendar. None of them could move the meeting. The gap was not intelligence — it was that nothing was allowed to act, so the last, tedious step always came back to me: open the other tab, copy the detail across, send the message.

Avagenc started from the opposite end. Assume the model can already reason well enough, and spend the engineering on the part that is actually hard — letting it touch real accounts without becoming something you have to supervise.

A single tool call is easy. The first time a request needed three in a row — read the thread, find the free slot, send the invite — the failure modes stopped being about the model at all. A calendar call would time out after the email had already been read, and the run would end somewhere in the middle with no way to say what had and had not happened.

That is what pushed every integration behind one interface and every irreversible step behind a confirmation. Gmail, Calendar, Contacts, Spotify and Tuya disagree about almost everything; the orchestrator should not have to know that any of them exist.

A Go orchestration layer, a Next.js web client and an Android client, containerised so the thing I run locally is the thing that runs in the cloud. Live integrations across five services, with conversation memory in a vector store so the assistant is still useful on the second day.

It is not finished. It is in the state where the interesting problems left are product problems rather than plumbing ones — which is where I wanted it to be.