OpenAI Codex launches subagents and custom agent support in general availability
OpenAI Codex subagents reached general availability after weeks of preview, enabling developers to define custom agents as TOML files and parallelize task execution. The feature mirrors Claude Code's implementation with default subagents for exploration, worker, and default operations.
OpenAI Codex subagents are now in general availability, moving out of preview status after several weeks of testing behind a feature flag.
The implementation follows Claude Code's model, offering three default subagent types: "explorer," "worker," and "default." The distinction between worker and default subagents remains unclear in OpenAI's documentation, though the worker variant appears designed for parallelizing large numbers of small tasks based on CSV handling examples.
Custom Agent Configuration
Developers can now define custom agents as TOML configuration files stored in ~/.codex/agents/. These custom agents support:
- Custom instructions tailored to specific use cases
- Model assignment, including gpt-5.3-codex-spark for high-speed execution
- Named references for direct invocation in prompts
OpenAI's documentation includes this example demonstrating the pattern:
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.
This prompt instantiates three named custom agents with specific responsibilities, each potentially running in parallel.
Industry-Wide Adoption Pattern
Subagent architecture has become standard across coding platforms. Competitors now support similar patterns:
- Claude Code: subagents with custom definitions
- Gemini CLI: experimental subagent support
- Mistral Vibe: agent selection and skills framework
- Visual Studio Code Copilot: built-in subagent layer
- Cursor: dedicated subagent documentation
- OpenCode: integrated agent system
What this means
Subagents represent the consolidation of agentic patterns in coding workflows. Rather than creating monolithic agents, developers can decompose complex tasks into specialized subagents with distinct capabilities and models. This approach reduces hallucination risk through task specialization and enables genuine parallelization for I/O-bound operations. OpenAI's integration of gpt-5.3-codex-spark into custom agents gives users explicit control over speed-accuracy tradeoffs per agent type.
Related Articles
GitHub Argues Chat Interfaces Fall Short for Coding Tasks, Points to Canvas UI as Alternative
GitHub's engineering blog makes the case that chat boxes aren't the right interface for every developer task, pointing to canvas-style UIs as a more tangible alternative within Copilot. The post is a design philosophy piece rather than a new model or feature launch.
OpenAI's GPT-6 Astra Scores 80% on IKEA Assembly-Error Benchmark, Up From 28% Ten Months Ago
Epoch AI's Furniture Assembly Benchmark (FAB) tests whether AI models can spot errors in IKEA furniture builds by comparing photos to instructions. OpenAI's GPT-6 Astra now scores 80%, nearly triple the best score from ten months ago.
Cline Desktop v0.0.35 Adds Linux Support, Fixes Plugin Slash Commands and Voice Input
Cline's desktop coding assistant now runs on Linux with native .deb and .rpm packages. The release also fixes plugin slash commands, restores live voice transcription, and refreshes the model catalog to 6,386 entries with Claude Opus 5.5 becoming the default for 11 providers.
Meta Opens Early Access Signups for New Muse AI Features Via In-App Prompt
Meta is letting users request early access to new Muse AI features by prompting the assistant directly, rather than running a traditional randomized beta test. The features, teased at Connect 2026, include a video-chat avatar, expanded shopping connectors, Mac computer-use capabilities, and support on Meta's AI glasses.
Comments
Loading...