reference codex 0 144 multi agent stable 20260715T141020

Verified directly from the Codex 0.144.4 binary's feature flags (not docs, actual strings/flags in the installed binary):

multi_agent            stable             true      <- enabled by default
multi_agent_v2         under development  false
enable_fanout          under development  false
use_agent_identity      under development  false

The binary carries real subagent machinery: spawn_agents_on_csv, x-openai-subagent, SubagentStartHook. agents.max_threads defaults to 6, agents.max_depth defaults to 1 (lead can spawn children, children cannot recurse). Subagents inherit the current sandbox policy by default, so under a full-access profile every spawned subagent is also full-access unless custom agents are defined with per-agent sandbox_mode = "read-only" in ~/.codex/agents/*.toml (project-scoped, e.g. ~/creatortrack/.codex/agents/). learn.chatgpt.com is a real OpenAI docs domain for this feature, not fabricated.

Why: this corrects an earlier wrong claim (made in-session) that Codex was single-agent with no fan-out; Codex's own claim about its subagent capability was accurate. How to apply: when advising Justin on Codex-based parallel workflows (e.g. lead + N auditor/reviewer subagents), assume multi_agent works natively in 0.144.4+; recommend read-only custom agent profiles for any subagent that should not have write/merge/deploy power, since sandbox inheritance is full-access by default.

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