If AI agents can be seen as AI assistants that perform tasks for users, then subagents are essentially 'smaller AIs' created to handle individual parts of a larger task. This idea is gradually becoming a trend in modern AI programming tools such as OpenAI Codex, Anthropic Claude Code, and Google Gemini CLI.
What exactly are Subagents?
In the simplest terms, subagents help break down a large task into smaller parts.
For example, a developer might ask an AI to read the codebase, write new features, run tests, check for bugs, and then compile the final results. If an AI has to handle the entire process simultaneously, the system can easily become chaotic or lose context at certain steps.
Subagents solve this problem by breaking down the work into multiple separate roles. Each smaller agent handles a specific part and then sends the results back to the main agent for aggregation.
In OpenAI's documentation, the Codex describes how it can create specialized subagents that run in parallel and then combine their results into a complete response. OpenAI suggests that this approach is particularly useful for multi-step tasks such as codebase exploration or the implementation of complex features.
Simply put, subagents are like dividing a large project among multiple team members instead of having one person handle everything.
Why are subagents becoming increasingly important?
Practical programming is rarely just about 'writing code'. A complete workflow typically involves many steps such as reading files, understanding the project structure, running commands, debugging, testing, and checking the final results. When AI has to handle this entire process in a single thread, the context can easily become confusing or inconsistent. Breaking down tasks across multiple subagents makes the workflow significantly clearer and easier to manage.
Interestingly, not all companies use the same terminology. Some call them 'subagents,' others 'agent teams,' and some simply describe them as AI agents capable of using tools. But the general idea remains quite similar: break down tasks so the system can process them more efficiently.
OpenAI Codex: the clearest example of subagents
Of the three tools mentioned, the Codex is perhaps the product most directly associated with the concept of subagents.
According to OpenAI, Codex is designed for multi-agent workflows and allows multiple agents to work in parallel on different projects. The system also supports built-in worktrees and cloud environments so that multiple agents can work simultaneously.
The key point is that OpenAI not only 'suggests' this capability, but also directly uses the term subagent workflows in its official documentation. This makes the Codex the clearest example when discussing subagents in AI coding today.
In other words, if anyone wants to understand how subagents actually work, the Codex is probably the easiest place to visualize it at the moment.
Claude Code: supports multi-agent but uses a different calling method.
Anthropic's Claude Code shares many similarities with the Codex, but the description differs slightly. According to Anthropic, Claude Code can read codebases, edit files, run commands, and manage projects directly from the terminal. The system also supports creating multiple Claude Code agents working in parallel on different parts of the same task.
Anthropic describes this model as a single lead agent coordinating, with multiple smaller agents handling different tasks before merging the results. In concept, this is quite similar to subagents. However, Anthropic doesn't use the term 'subagents' directly in their overview page. Therefore, the most accurate description is probably: Claude Code supports the 'agent teams' model, while the Codex is more directly associated with the concept of subagents.
Gemini CLI: A terminal-first and open-source approach.
Google's Gemini CLI takes a slightly different approach. Google describes Gemini CLI as an open-source AI agent for the terminal, allowing users to access Gemini directly from the command line. The system operates on a 'reason-and-act loop' model, combining built-in tools and a local or remote MCP server to handle tasks such as bug fixing, feature creation, or test coverage improvement.
Compared to Codex or Claude Code, Gemini CLI leans more towards a terminal-first experience and the ability for developers to directly control the workflow.
However, based on current documentation, Google primarily describes Gemini CLI as an AI agent rather than a specific subagents system. Therefore, it's more appropriate to view it as part of the overall trend of AI agent coding tools.
AI coding tools are gradually becoming more like a 'team' than a chatbot.
Interestingly, all three tools are heading in the same direction: transforming AI from a chatbot that answers questions into a system capable of independently dividing tasks and handling more complex workflows. This is crucial because developers now want AI to not just generate code. They want AI to understand projects, track long contexts, run multiple steps consecutively, and coordinate work autonomously.
And that's where subagents, or multi-agent workflows, start to become much more useful than the traditional chatbot model.
If we were to summarize the concept of subagents in the simplest way, we could understand them as smaller AIs created to handle individual parts of a larger task.
Of the three tools compared, the Codex is the clearest example of subagents because OpenAI directly uses and describes this concept. Claude Code also supports multi-agent workflows but uses a different terminology, while Gemini CLI is a terminal-first AI agent focused on workflow and developer control.
Despite their different approaches, all three demonstrate a clear trend: AI coding tools are gradually shifting from 'chat assistants' to systems capable of coordinating complex tasks like a real team.
You've just finished reading the article "What are Subagents? A comparison of OpenAI Codex, Claude Code, and Gemini CLI." edited by the TipsMake team. We hope this article has provided you with many useful tech tips and tricks. You can search for similar articles on tips and guides. Thank you for reading and for following us regularly.