Summary
- • OpenAI Codex subagents move from feature-flag preview to general availability
- • Default agents ship with 'explorer', 'worker', and 'default' roles out of the box
- • Custom agents defined as TOML files can be assigned to specific models including gpt-5.3-codex-spark
- • Multi-agent architecture is now a standard pattern across major AI coding platforms
Details
Codex subagents reach general availability after weeks in feature-flag preview
The feature graduated from limited preview to full GA, making multi-agent coordination available to all Codex users. The preview period allowed OpenAI to validate stability and behavior before broad rollout.
Three default subagents ship out of the box: explorer, worker, and default
Default roles enable immediate use without custom configuration. Based on documentation examples, 'worker' appears designed for running many small parallel tasks at scale, while 'explorer' likely handles investigation-oriented subtasks. The distinction between 'worker' and 'default' is not fully documented.
Custom agents defined as TOML files in ~/.codex/agents/ with per-agent model assignments
Developers can specify custom instructions and assign specific models to each agent, including gpt-5.3-codex-spark for latency-sensitive workloads. Agents are then referenced by name in natural language prompts, enabling clean orchestration syntax.
Documentation example shows multi-agent coordination across debug, map, and fix roles
The sample prompt assigns specialized named agents to distinct subtasks within a single instruction — reproducing a bug, tracing the code path, and implementing a minimal fix — demonstrating how orchestration decomposes complex engineering tasks.
Multi-agent architecture now broadly adopted across major coding agent platforms
Codex's implementation closely mirrors Claude Code's subagent pattern. Multi-agent coordination has become a standard architectural pattern in AI-assisted coding tools rather than a differentiator, suggesting rapid convergence on shared design norms.
gpt-5.3-codex-spark available as a speed-optimized model option for custom agents
The inclusion of a named speed-optimized model variant suggests OpenAI is offering tiered model options within Codex, letting teams trade capability depth for throughput depending on the agent's role in a larger workflow.
Product Launch = GA release milestone; New Tech = new capability details; Tech Info = implementation examples; Industry Update = ecosystem convergence; Context = background on available models
What This Means
OpenAI's GA release of Codex subagents formalizes multi-agent coordination as a first-class feature in its coding platform, matching what Claude Code already offers. The ability to define custom agents with specific model assignments gives development teams fine-grained control over how AI breaks down and executes complex engineering tasks. As the same architecture spreads across platforms, multi-agent coding workflows are transitioning from experimental to expected — developers building on any major AI coding tool will increasingly encounter and need to design for this pattern.
Sources
- Use subagents and custom agents in CodexSimonwillison