Features
- Actor-per-agent pattern with a coordinating manager Rivet Actor
- Queue-based intake using
c.queue.nextinside the run loop - Sandbox Agent SDK sessions per agent with streamed output
- Persistent history stored in Rivet Actor state
- Live status updates via events and polling
Prerequisites
- Sandbox Agent endpoint configured with
SANDBOX_AGENT_URLandSANDBOX_AGENT_TOKENif you are not running locally - API key for the coding agent you choose inside the Sandbox Agent environment, such as
OPENAI_API_KEYforcodex
Implementation
Each AI agent Rivet Actor creates or reuses a Sandbox Agent session, streams turn events, and broadcasts deltas back to the UI.
- Sandbox agent integration (
src/actors.ts) - Frontend orchestration (
frontend/App.tsx) - Server entry point (
src/server.ts)
Resources
Read more about queues, run handlers, state, and events.


