reference codex multi agent stable default 20260715T135235

Codex CLI (installed version 0.144.4) ships real multi-agent subagent orchestration, not a single-agent tool. Confirmed by inspecting the binary's feature flags directly: multi_agent is stable and enabled by default, multi_agent_v2/enable_fanout/use_agent_identity are still under development and off. The binary carries real machinery: spawn_agents_on_csv, x-openai-subagent, SubagentStartHook. Docs live at learn.chatgpt.com/docs/agent-configuration/subagents (verified genuine, not fabricated, despite initially looking like a bogus domain).

Key mechanics: subagents inherit the current sandbox policy (so under a full-access profile like devfull, every spawned subagent is also full-access, unless scoped down via custom agent files with sandbox_mode = "read-only" in ~/.codex/agents/<name>.toml). Defaults: agents.max_threads = 6 concurrent, agents.max_depth = 1 (children can't recurse further).

Why: this corrects a prior wrong assumption (in reference_codex_workspace_write_git_readonly.md) that Codex CLI was single-agent with no parallel fan-out. That claim is now outdated/wrong for current Codex releases.

How to apply: when advising on Codex CLI capabilities or writing Codex prompts/AGENTS.md instructions that reference parallel review/audit patterns, assume subagent fan-out works natively. If auditors/reviewers need to be prevented from writing (e.g. under a full-access profile), scope them via project .codex/agents/*.toml with sandbox_mode = "read-only" rather than assuming the base profile protects against writes.

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