reference codex multi agent stable default 20260715T135428

Codex CLI (0.144.4, codex aliased to codex --profile devfull on Console) has genuine multi-agent subagent orchestration, verified against the binary's feature flags: multi_agent is stable and true by default; multi_agent_v2, enable_fanout, use_agent_identity are still under development/false. The binary contains real machinery (spawn_agents_on_csv, x-openai-subagent, SubagentStartHook). Custom per-role agents can be defined in ~/.codex/agents/*.toml (or project-scoped <repo>/.codex/agents/) with per-agent sandbox_mode (e.g. read-only for auditors) even when the parent session runs under a full-access profile like devfull — subagents inherit the parent's sandbox by default unless overridden. Defaults: agents.max_threads=6, agents.max_depth=1 (no recursive spawning). Docs at learn.chatgpt.com/docs/agent-configuration/subagents are real (verified, not fabricated) despite looking like a suspicious domain at first glance.

Why: I initially told Justin Codex was single-agent with no parallel fan-out, which was wrong and outdated; I also wrongly called the docs URL fabricated before checking. Both corrected against the actual installed binary.

How to apply: don't assume Codex CLI lacks subagent/parallel capabilities; check the installed binary's feature flags before asserting capability claims about Codex. When scoping Codex subagents for audit/review roles, use per-agent sandbox_mode = "read-only" config files rather than relying on the parent profile's access level.

[auto-memory session 577603bc-d3a9-4b19-bb8e-6db80243b054, confidence 0.75, mode staged]