Claude Code, MCP, Permissions, Hooks, Subagents, Skills, and Memory

Reflection – Week 2

Yinusa Kolawole Gbenga

Week 2 of the DevOps Micro Internship (DMI) Cohort 3 with Agentic AI was a major shift in how I think about DevOps automation and AI-assisted development. Rather than simply using AI to generate commands or answer questions, I started learning how to build a more structured working relationship with an AI coding agent using Claude Code.

One of the biggest areas I explored this week was Claude Code itself. I learned how it can work directly inside a project, inspect files, execute commands, understand project instructions, and help with development tasks. What stood out to me was that effective use of an agent requires more than simply giving it a request. I need to understand the project structure, provide clear instructions, review what the agent does, and verify the results.

I also worked with MCP (Model Context Protocol) and learned how external tools can be connected to Claude Code. This helped me understand how an AI agent can move beyond the files in a project and interact with additional tools and services. I encountered configuration issues with environment variables while setting up GitHub MCP. That experience reinforced an important DevOps lesson: configuration and secrets management matter. I learned not to place sensitive tokens directly inside project configuration files and instead use environment variables.

Another area that changed my perspective was Memory. I experimented with project-specific memory and tested whether Claude Code could recall a project convention after the previous session had been closed. Claude correctly remembered the project's no-JavaScript rule in a fresh session. I also learned that an AI agent should not blindly store every instruction as fact. When I asked it to remember a CSS gradient that did not actually exist in the project, it checked the code first and declined to store the incorrect information. That was a useful lesson about verification and trust.

I also gained a better understanding of Permissions, Hooks, Skills, and Subagents as mechanisms for controlling and organizing agentic workflows. These concepts helped me see that agentic AI is not simply about asking an AI to do everything. It is about creating boundaries, reusable capabilities, automation points, and specialized workflows that make the agent more predictable and useful.

The biggest challenge for me this week was getting comfortable with the amount of new terminology and configuration involved. There were moments when a seemingly small configuration issue required several checks before I understood what was happening. Instead of seeing these problems as failures, I am beginning to treat them as part of the learning process.

My biggest takeaway from Week 2 is that I should not blindly trust automation. I need to inspect, verify, understand, and then accept the result. An AI agent can make development faster, but I still need the technical understanding to judge whether what it is doing is correct.

My Habit Going Forward

One habit I plan to implement is a verify-before-accept workflow:

Ask → Observe → Verify → Understand → Accept → Document

Whenever I use an AI agent to make a change, I want to check the files, commands, configuration, and output before considering the task complete. I also want to document important lessons and project conventions so that they can be reused in future sessions.

Week 2 has made me more interested in the possibilities of Agentic AI within DevOps. I am still learning, but I now have a clearer understanding that becoming effective with these tools is not about giving AI complete control. It is about learning how to work with an agent responsibly, securely, and systematically.

Yinusa Kolawole Gbenga