Codex Launches Subagents and Custom Agents, Expanding AI Workflow Flexibility

OpenAI’s Codex platform now ships subagent workflows as a general‑availability feature, moving the capability out of the preview flag that developers have been testing for the past several weeks. The update lets a single Codex request spawn multiple specialized agents that run in parallel, each with its own model configuration and toolset, before Codex aggregates their outputs into a single response — a pattern that mirrors the “explorer,” “worker,” and “default” agents introduced in Claude Code and other emerging coding‑agent ecosystems [Simonwillison].

The practical upside is most evident in tasks that naturally decompose into independent subtasks. For example, a developer can ask Codex to review a pull request across six dimensions—security, code quality, bugs, race conditions, test flakiness, and maintainability—by issuing a prompt that spawns one subagent per point, waits for all six analyses, and then returns a concise summary for each [OpenAI]. Codex orchestrates the entire lifecycle: it creates the subagents, routes follow‑up instructions, monitors completion, and finally consolidates the results. The CLI exposes a /agent command that lets users switch between active threads and inspect ongoing work, while the Codex app already surfaces subagent activity; IDE extension support is slated for a later release [OpenAI].

Custom agents are defined via TOML files placed in ~/.codex/agents/, where developers can prescribe bespoke instructions and bind each agent to a specific model. This flexibility enables use cases such as “Investigate why the settings modal fails to save. Have browser_debugger reproduce it, code_mapper trace the responsible code path, and ui_fixer implement the smallest fix once the failure mode is clear,” a workflow demonstrated in the official documentation [Simonwillison]. By allowing model‑selection guidance at the agent level, Codex mitigates classic agentic pitfalls like context pollution and context rot, though the trade‑off is higher token consumption because each subagent runs its own model inference [OpenAI].

The move aligns Codex with a broader industry trend toward “agentic engineering,” where swarms of specialized AI agents collaborate on complex software tasks. TechCrunch notes that the rise of agentic coding “is already having a seismic impact on how software is written,” with subagents handling much of the grunt work that developers traditionally performed [TechCrunch]. ZDNet’s coverage of OpenAI’s Spark model—an iteration that reportedly codes 15× faster than GPT‑5.3‑Codex—underscores the performance pressure driving these architectural innovations, even as the faster model introduces its own set of constraints [ZDNet].

Analysts and early adopters caution that while subagents broaden Codex’s applicability, they also introduce new operational considerations. Parallel execution can exacerbate token costs and latency, especially when many agents are instantiated simultaneously. Moreover, the distinction between “worker” and “default” agents remains opaque; Simonwillison speculates that “worker” agents are intended for “running large numbers of small tasks in parallel,” but official clarification is pending [Simonwillison]. As OpenAI rolls out IDE integration and refines the custom‑agent API, developers will gain clearer visibility into these trade‑offs, shaping how subagent workflows are adopted in production pipelines.

In sum, Codex’s subagent and custom‑agent capabilities mark a decisive step toward more modular, parallelizable AI‑assisted development. By exposing orchestration logic, model selection, and tool integration under a unified interface, OpenAI gives engineers the building blocks to construct sophisticated, multi‑agent pipelines without writing bespoke orchestration code. The feature’s general availability, combined with upcoming IDE support, positions Codex to compete directly with parallel‑agent implementations in Claude Code, Gemini CLI, Mistral Vibe, and other emerging platforms [Simonwillison]. As the ecosystem coalesces around agentic workflows, the next frontier will be measuring productivity gains against the higher token overhead and ensuring that the added flexibility translates into tangible reductions in development cycle time.