Hacking Subagents Into Codex CLI — Brian John, Betterup

Brian John, a Principal Full-Stack Engineer at BetterUp, discusses how to implement sub-agents in the Codex CLI. He explains the benefits of using sub-agents for context management and shares his design for hacking sub-agents into Codex CLI, which involves a parent Codex session running a wrapper script to initiate a child Codex instance as the sub-agent. He addresses the challenges he faced with Codex's sandbox permissions and details the minimum required permissions for both the parent and child processes. Brian also touches on security considerations, referencing Meta's "Agent's Rule of Two," and provides a proof-of-concept repository with toy agents, a wrapper script, and a Codex wrapper script. He demonstrates how to use the word counter and file writer sub-agents within Codex, highlighting the need to grant permissions and the slower, serial execution compared to tools like Claude.