Practical conventions for running several coding-agent sessions on one machine at once without them colliding or becoming hard to track:
Use these together when running more than one or two agents locally: name the session, reserve its port range up front, and lean on a port-manager tool like Portless once the number of concurrent work-trees makes manual tracking error-prone.
Foreground / Background / Remote Agent Execution Model
Rather than watching one agent iterate live, fire off multiple coding agents on different tasks simultaneously and check back on each between other work (e.g., before a meeting, then reviewing all of them after). This round-robin pattern trades continuous supervision for periodic steering across several agents at once.
Ryan Nystrom (Notion) runs many git work trees at once, each with its own local dev server on a sequential port (3000, 3001, 3002 ... 3009), so several concurrent Codex sessions on the same repo don't collide on ports or working-directory state. He separately uses Codex's own "projects" feature to group work by codebase/initiative — a lighter-weight organizing layer on top of, not a replacement for, the work-tree-per-task setup, useful for keeping context and history separated when juggling many different projects at once.