A system that can read itself
An assistant I text on WhatsApp. Ten services underneath it, and an instrumentation layer that has found more than testing did.
BenAri
01 — Built to be used
34
Tools, six of them locked
Calendar, email, tasks, notes. Anything that sends or deletes stops and waits for a typed yes — a gate in the code, not an instruction in the prompt, so the model cannot talk its way past it.
71
Named events, every turn measured
Every turn is written back as a trace: what it recalled, what it called, how long each stage took. That layer has caught three problems before they ever reached me — none of which a test would have found.
4
Tools pointed at its own source
It can open a file in its own repo, search the tree, and read its own commit log — read-only, sandboxed. Ask why it behaves a certain way and it goes and looks.
6×
Cheaper per turn than it looks
A turn re-sends its persona, its rules and its whole tool list up to six times as the model works. The prompt is split so every one of those after the first is served from cache.
Seven services, running since June
Five background agents keep it current between conversations — one of them measures how often it gets a memory wrong and tunes its own threshold on the result. Voice is live end to end. A new domain is a row in a table, not a rewrite.