Follow up to my Codex beginner guide. The six terms I told you to wait on, explained in plain language. AGENTS.md, Memory, Skills, Subagents, Hooks and MCP.
By Holidays in Europe / April 27, 2026 / No Comments / Uncategorized
Understanding Key Concepts in Codex: A Clear Guide for Beginners
As a continuation of my previous beginner’s guide to Codex, I wanted to take a moment to clarify six essential terms that often come up when working with this powerful AI tool. My goal is to explain these concepts in straightforward language, avoiding technical jargon and complex terminology, so you can understand and utilize them effectively. Let’s dive into each term and explore how they fit into your workflow with Codex.
- What is AGENTS.MD?
AGENTS.MD is a simple markdown file that Codex reads every time you start a new session. Placed in your home folder or within your project directory, it requires no setup or configuration—just drop the file into the right place. This file serves as a set of instructions or guidelines for Codex, helping it understand how you want it to behave.
In your AGENTS.MD file, you might include a brief summary of what you do, how you prefer Codex to communicate, or any important rules—such as asking for permission before deleting files. Keep it concise and written as if you were speaking to a person, not programming software. For example, if you’re managing a small accounting firm, you might specify a formal tone for client interactions and instructions like “ask before executing shell commands.” You can refine these instructions over time as you observe how Codex responds.
- Understanding Memory in Codex
Memory in Codex refers to its ability to remember context from previous interactions. By default, this feature is turned off to prioritize privacy and security, but you can enable it if you want Codex to carry information across different sessions automatically.
When activated, Memory stores notes in your Codex home folder, allowing the AI to recall pertinent details without you needing to reintroduce them. To turn it on, simply enable the setting in your configuration options or add a specific line to your setup file. Note that in regions like the European Economic Area, the UK, and Switzerland, this feature was not initially available at launch, so verify whether it’s accessible for your account.
Important tip: Avoid storing sensitive information such as client data or confidential account details in Memory. Its purpose is to maintain context for ongoing tasks, not to serve as a secure vault.
- Building Skills in Codex
A skill is a small, organized collection of instructions and resources for executing a specific, repeatable task. It typically consists of a folder containing a file named SKILL.md, which describes the workflow you want Codex to perform regularly. Additional scripts or resources can be included if needed.
For instance, you might create a skill for generating weekly status reports from your emails, formatting meeting notes, or cleaning up PDFs. Once established, Codex can recognize and invoke the appropriate skill automatically based on the topic, or you can tell it explicitly to use it.
Developing skills helps automate routine activities and saves time. It’s also worthwhile to explore existing skills or plugins shared by others before creating your own, as many common workflows have already been packaged.
A useful tip: Your first skill should target a task you find yourself explaining or performing repeatedly—when you notice you’re doing the same thing multiple times, that’s the perfect candidate to turn into a skill.
- Using Subagents for Parallel Tasks
Subagents are auxiliary “workers” that you can delegate specific parts of a task to, allowing multiple activities to happen simultaneously. Codex only spawns subagents when you explicitly request parallel processing; it does not generate them automatically.
This feature is especially handy when tasks are independent of each other, such as researching multiple documents at once, comparing several contracts, or analyzing different lists against specific criteria. Codex offers built-in subagents like the default worker for general tasks, a heavier-duty worker for complex execution, and an explorer for digging through code or files.
Note: Since each subagent consumes tokens (the credits used for processing), it’s best to limit the number of subagents to what is necessary. Using too many can slow down your workflow.
Most tasks are completed more efficiently through focused, sequential work. Use subagents judiciously to optimize your productivity.
- Hooks: Automating Actions with Triggers
Hooks are small, automatic triggers that activate when certain events occur during your session—such as starting a new session, before Codex calls a tool, or after a file has been modified. They enable you to run custom checks or notifications at precise moments, helping streamline your workflow.
Currently, this feature is flagged in Codex’s settings and requires enabling in your configuration. While hooks can automate repetitive actions, their implementation is still evolving, and they may not be entirely reliable for critical tasks.
A practical tip: If you’re new to Codex or don’t yet need automation, you can skip hooks for now. Revisit this feature once you find yourself repeatedly performing the same small task at the same point in your workflow.
- MCP: Integrating External Services
MCP (short for “Module Communication Protocol” or similar, depending on context) is how Codex connects with external tools and services. Examples include GitHub, Figma, documentation servers, databases, or custom internal tools.
To set up an MCP connection, you add a configuration block with the necessary connection details. Once configured, Codex can access these services seamlessly during your sessions. The process typically involves logging into the external service once via your browser, after which subsequent connections are automatic.
An important advantage: MCP is not exclusive to Codex. The same connectors can often be used with other AI systems, reducing vendor lock-in.
A tip to get started: Focus on integrating one service you currently need to improve your workflow. As your tool set grows, keep in mind that adding too many integrations can complicate your setup, so aim for a lean and efficient configuration.
Final Thoughts
These six concepts—AGENTS.MD, Memory, Skills, Subagents, Hooks, and MCP—are foundational to customizing and optimizing your experience with Codex. By understanding and applying them thoughtfully, you can automate routine tasks, maintain better context, and streamline your overall workflow.
If you would like further details on any of these topics or have additional questions, feel free to comment below or reach out via direct message. Happy Codexing!