Product Release And Access Change

Issue 75 Edition 2026-03-16 4 min read

Not accepted General

Sources: 1 • Confidence: Medium • Updated: 2026-04-13 03:49

Key takeaways

OpenAI Codex subagents are generally available after several weeks of preview behind a feature flag.
In Codex, custom agents can include custom instructions and can be pinned to specific models, including gpt-5.3-codex-spark.
In Codex, custom agents can be referenced by name in prompts to orchestrate multi-step workflows where different agents reproduce bugs, trace code paths, and implement fixes.
Available information does not clearly explain the distinction between the worker Codex subagent and the default Codex subagent.
The subagents pattern is supported across multiple coding-agent platforms including Codex, Claude Code, Gemini CLI, Mistral Vibe, OpenCode, Visual Studio Code, and Cursor.

OpenAI Codex subagents are generally available after several weeks of preview behind a feature flag.

In Codex, custom agents can include custom instructions and can be pinned to specific models, including gpt-5.3-codex-spark.

In Codex, custom agents can be referenced by name in prompts to orchestrate multi-step workflows where different agents reproduce bugs, trace code paths, and implement fixes.

Available information does not clearly explain the distinction between the worker Codex subagent and the default Codex subagent.

The subagents pattern is supported across multiple coding-agent platforms including Codex, Claude Code, Gemini CLI, Mistral Vibe, OpenCode, Visual Studio Code, and Cursor.

What is the precise behavioral difference between a worker subagent and a default subagent in Codex (capabilities, permissions, tool access, lifecycle, routing)?
What are the operational limits for Codex subagents (max concurrent subagents, latency, timeouts, context-sharing model, and failure/retry behavior)?
How does model pinning interact with pricing and quotas (per-agent billing, per-model rate limits, and whether gpt-5.3-codex-spark has distinct constraints)?
Do named custom agents in Codex share memory/state across invocations, and if so what are the boundaries (project, repo, session, or account)?
What minimum common functionality qualifies as “subagents pattern support” across the listed platforms (configuration format, invocation syntax, tool permissions, and interoperability)?

General availability of Codex subagents could expand adoption of multi agent coding workflows, increasing usage of coding agent platforms that implement named delegation and agent configuration.
Per agent model pinning including gpt-5.3-codex-spark could enable deliberate mixing of models by task, potentially shifting usage patterns toward specialized models within the same workflow.
Cross platform mention of subagents pattern support suggests convergence on a common workflow primitive, potentially reducing differentiation and making interoperability and ecosystem tooling more important.

Clear documentation of worker versus default subagent roles including capabilities, tool access, lifecycle, and routing, enabling teams to standardize configurations.
Published operational limits for subagents including concurrency, latency, timeouts, context sharing, and retry behavior, indicating production readiness and predictable performance.
Pricing and quota details for model pinned agents including whether billing and rate limits vary by model such as gpt-5.3-codex-spark, clarifying economic incentives for adoption.

Continued ambiguity around subagent role taxonomy and permissions, preventing reliable configuration and limiting enterprise or team wide rollout.
Restrictive or unstable operational limits such as low concurrency or frequent timeouts, reducing the practicality of orchestrated multi step workflows.
Model pinning introduces unclear or unfavorable pricing and quota constraints, discouraging use of specialized agents and limiting workflow mixing.