In a move that could redefine AI-assisted coding and project management, OpenAI’s Codex has quietly rolled out a suite of groundbreaking features, including subagent swarms, collaboration modes, and new tools designed to enhance hierarchical agent orchestration and planning. The updates, spotted by developers on social media, signal a major leap forward in how AI agents collaborate and execute complex tasks.

The excitement began bubbling on X (formerly Twitter) when user @LLMJunky shared insights into the activation of subagents within Codex. “Oh snap! Subagents are actually live right now ‘officially’ in Codex. Sneaky!” posted @LLMJunky, advising users to enable the feature by setting “collab = true” in their config.toml file. The post, which garnered hundreds of likes and sparked discussions among developers, highlighted the potential for more configurable options in upcoming weeks.

Building on this, a follow-up thread from the same user speculated on an impending major release. “I suspect a major Codex release is on the horizon, including native Plan Mode,” they wrote, pointing to newly added prompts on OpenAI’s GitHub repository. These include support for subagents, a “Pair Programming” collaboration mode, Plan Mode, and a new “request_user_input” tool—also known as AskUserTool—that allows the AI to seek clarifying questions from users via an accept/deny interface.

At the heart of these enhancements is the Orchestration Agent, dubbed “Swarm Mode,” which acts as a conductor for subagents. This enables hierarchical structures where subagents can spawn additional subagents, all sharing access to the same skills, prompts, and tools. “Subagents can spawn other subagents can spawn other subagents,” explained @LLMJunky, emphasizing the system’s ability to handle unblocked tasks in parallel loops. Developers are encouraged to mark tasks as complete and log changes, ensuring seamless workflow progression.

Screenshots shared in the posts illustrate the feature in action. One depicts a Codex session where a subagent is spawned to explore a repository structure, with prompts like “Explore repo structure, key files, summarize relevant areas. Use ck - help first for search per instructions.” The subagent reports back with a concise map, demonstrating efficient delegation without cluttering the main context window.

Community reactions have been overwhelmingly positive, with users praising the productivity boost. “Productivity in Codex just went parabolic. Subagent Swarms. It’s a thing of beauty, really,” commented @LLMJunky in a related post. Others inquired about use cases, such as improving model performance or token efficiency. “It’s really about speed. And sometimes it can actually save tokens,” responded the poster, noting applications like repo searches or web research where subagents handle grunt work and return only essential outputs.

Plan Mode appears to be already accessible under experimental flags. Users can activate it with commands like “Switch to plan mode” or “Pair Programming Mode” after enabling collaboration features and updating to version 0.88 via /experimental. This mode allows for detailed planning before execution, with the AI breaking down tasks and seeking user input as needed.

The updates draw from prompts available on GitHub at https://github.com/openai/codex/tree/main/codex-rs/core/templates, where developers can dive into the raw implementations. As one commenter noted, “When spawning multiple agents, you must tell them they are not alone”—a quirky reminder of the nuanced prompting required to optimize these systems.

While OpenAI has yet to make an official announcement, the community’s testing suggests these features could transform Codex from a solo coding assistant into a full-fledged AI orchestra for software development. With calls for “BATMAN ACTIVATED” and “Garlic” (possibly hinting at future integrations), the developer ecosystem is abuzz with anticipation.

As AI tools continue to evolve, Codex’s embrace of swarm intelligence and collaborative workflows positions it at the forefront of autonomous agent technology. Developers eager to experiment should update their setups and explore the new configs—but be prepared for a paradigm shift in how projects are planned and executed.

View post on X

View post on X