Orchestrating a virtual engineering team over long-horizon tasks with Codex Autorunner

Problem

David Zhang wanted to spin up a virtual engineering team on any machine — a cloud box, a Mac Mini, a Linux server — and drive it from anywhere, including his phone. Existing orchestrators tie agents to a single machine, and plain 'goal mode' tends to wander rather than execute a long plan deterministically.

Approach

Codex Autorunner (open source) is a meta-harness that drives underlying coding agents (Codex, OpenCode, Hermes) through a series of tickets, curating each agent's prompt with context about past and future tickets plus a shared, prescriptive spec document. Because agents see what's done and what's pending, they can predictively adjust future tickets based on discoveries mid-implementation — David likens it to a 'Brownian bridge' that anchors iterations. It's controllable from a web UI and from Discord/Telegram (start on one device, continue on another), and a built-in project-manager agent delegates to and keeps context across coding agents. 'Services' publish preview links from the same URL, and cron automations can themselves write tickets. He self-hosts on Proxmox across three Linux servers with VMs reached over Tailscale, shipping pre-built VM images with Autorunner, and uses Hermes as the default PM (GPT-5.5 for high-stakes infra, GLM for cheap tasks).

Results

The tool has 73 releases since mid-January and wrote roughly 90% of itself. It runs long-horizon build plans (~12 hours) unattended: David front-loads deep planning into the spec + lean tickets and back-loads human QA/polish, and says he hasn't written a line of code in about eight months. The main friction is that long-running agents are genuinely slow to demo — a blessing and a curse of the approach.