Codex only spawns subagents when you ask it to. Use direct prompts:
Spawn security_auditor to review this authentication change for exploitable risks.
Spawn test_generator to add regression coverage for the bug fixed in this diff.
Use the narrowest specialist that can complete the job. Use orchestrator only when the request spans multiple domains or needs coordinated sequencing.
Read-only agents are for exploration, review, planning, research, and security analysis. Workspace-write agents can implement changes, but they still inherit the parent Codex approval policy and runtime overrides.
Include goal, relevant paths, constraints, and done criteria. For example:
Spawn backend_developer to add rate limiting to the login endpoint. Follow existing Express middleware patterns in src/middleware, preserve the API shape, and add the smallest relevant tests.
The project config uses max_depth = 1 to prevent recursive fan-out. For large work, ask the parent Codex session to spawn several named agents and then consolidate the result.