OpenAI Codex adds parallel subagent workflows

Codex Subagents moves agentic coding from single-thread prompting to deliberate workflow orchestration, which is a meaningful upgrade for real engineering teams.

• –Teams can split multi-step work like PR review, bug triage, and codebase exploration across focused agents, then merge outputs into one response.
• –Built-in agent roles (`default`, `worker`, `explorer`) plus custom TOML-defined agents make specialization practical without rewriting the whole workflow.
• –OpenAI keeps delegation explicit (no automatic spawning), which improves control and predictability for production use.
• –The main tradeoff is cost: parallel subagents consume more tokens than single-agent runs, so teams need to use them selectively.