Factory: issue-to-PR workflows with Neo and Codex

The speaker demonstrates Factory, an open source Rust project that coordinates an AI agent workflow around GitHub issues. It runs on a local computer or a virtual machine and polls for labels that trigger work. The walkthrough focuses on development automation rather than running model inference locally.

A manual example uses the Neo coding assistant to take a ticket through planning, an isolated Git worktree, implementation, tests and review before handing a pull request to a human. The speaker says the same workflow can apply to Claude Code, Codex or Pi.

Factory then automates the handoff. A refinement label triggers Codex to inspect a vague bug report and turn it into an implementation-ready ticket. After reviewing that ticket, the speaker adds an implementation label. A separate bug-finder workflow can run on a schedule or manually and open an issue for later review.

The speaker uses Git worktrees locally and discusses Docker sandboxes for VM or team use, while acknowledging that the sandboxing approach is still unsettled. They describe Factory as early-stage software and say deterministic polling consumes no model tokens until an agent has work. Agent execution still costs tokens, and the speaker reserves some tasks for interactive work and human judgment.