Running dozens of coding agents in parallel requires moving up an abstraction layer: one orchestrator agent spawns and manages many headless sessions. Both Claude Code and Codex support this via CLI flags — `claude -p "prompt"` and `codex exec "prompt"` — which launch isolated sessions that execute a task autonomously and return only the final result. The orchestrator sees logs optionally and receives a concise summary such as "The task was fixed and merged to dev." Additional flags control permissions, max turns, and model selection, letting the orchestrator fire off hundreds of sessions without manual intervention.

Effectiveness hinges on three practices. First, every headless agent must verify its own work — tests, linters, or custom scripts — because the human cannot review each decision. Second, tasks must suit autonomous execution; refactoring is ideal. The author feeds a repository to Claude Fable (the newest flagship model) with a prompt to "look for refactor opportunities... in a prioritized manner in an HTML file and make a plan on how to fix it." The resulting plan then drives a second orchestrator prompt that spins up Claude Opus headless sessions for each independent fix, running them in parallel until the refactor completes.

Third, agents need full tool access: MCP servers, unrestricted file operations, and permission to run commands without prompting. The goal is a one-shot prompt that behaves like a supercharged sub-agent. The author reports daily use of 100+ concurrent sessions, with the orchestrator pattern cutting human oversight to near zero while pushing large-scale refactors to production.

This architecture shifts the bottleneck from human attention to agent coordination quality. As models gain reliability in self-verification and task decomposition, the orchestrator pattern becomes a force multiplier — but it demands rigorous verification hooks and clear task boundaries to avoid compounding errors across hundreds of parallel runs.