I’m building Super Hornet, the agent workspace for the AI you already use.

What’s hard about it

1. Runs that outlive every client. Runs used to stream from inside the web process, so a deploy killed them. Now a run is a row in Postgres. A worker claims it, owns the model calls and the tool loop, appends every event in order, and hands off at safe boundaries during a deploy. A device that reopens a running thread starts from the worker’s own progress row, so it is live at once instead of replaying the turn.
2. Four providers, one transcript. Each lane streams, reasons, and fails differently. Reasoning is replayed per provider, so a thread can change models mid-conversation. A reply that hits a rate limit settles its running tools, closes its open blocks, and continues on the next account without an error.
3. An agent in a person’s own browser. Commands reach the person’s active device as claim-once rows over Postgres and server-sent events. Passwords, codes, and card numbers are redacted at the source before the model sees a page. Purchases and sends confirm the way the person chose, and a touch on the page interrupts the agent’s action.
4. Containment. Every conversation runs in its own microVM. GitHub tokens are minted per operation, provider credentials are encrypted at rest with AES-256-GCM, and in Ask mode a command waits on an approval card that shows exactly what will run.