Reflection – Week 2
Week 2 of the DevOps Micro Internship pushed me from "using AI as a chatbot" to actually building with it. I spent most of my time on Skills, CLAUDE.md, Subagents, and MCP — and this week changed how I think about both AI and DevOps.
1. Biggest technical insight I got this week
My biggest insight was understanding how subagents and MCP work together. Before this week, I thought Claude Code was just one AI doing everything in a single conversation. Now I understand that work can be divided: subagents handle focused tasks with their own context and their own restricted tool access, while MCP connects the AI to real external systems instead of keeping it locked inside the terminal. Combined with skills, which turn long, repeated prompts into reusable slash commands, and CLAUDE.md, which acts like a standing set of rules the AI reads before every task, the whole setup started to feel less like "chatting with AI" and more like configuring a small, disciplined team.
2. Biggest insight I got about myself this week
I learned that I understand things much better by doing than by reading or watching. Concepts like MCP made very little sense to me in theory, but the moment I connected a server and watched the agent actually use it, everything clicked. I also realized that I tend to move fast and assume I know where things go — which directly connects to my biggest weakness below.
3. My biggest weakness or loop I noticed
My repeated loop this week showed up while setting up hooks. The relative path I mentioned in settings.json could not locate my bash script, and the hook kept failing with an error I initially didn't understand. I went around in circles checking the script itself, when the script was fine — the problem was the path configuration. With the help of AI, I understood that I needed to anchor the path to the project directory instead of relying on a relative path:
bash "$CLAUDE_PROJECT_DIR/.claude/hooks/user-prompt-guard.sh"
Once I used $CLAUDE_PROJECT_DIR, the hook resolved correctly. The real weakness this exposed wasn't the config itself — it was that I assumed the path would "just work" instead of verifying how it was actually being resolved.
4. One system I will implement from this week
The habit: Whenever I reference a file path in any configuration (like settings.json for hooks), I will verify it before running anything — echoing the resolved path, checking the file exists with ls, and preferring absolute or variable-anchored paths like $CLAUDE_PROJECT_DIR over bare relative paths.
When: Every time I write or edit a config file, before I test the workflow that depends on it.
How it will help: This removes my most common failure mode. Verifying a path takes 30 seconds; debugging a hook that silently fails because of a wrong path took me much longer. It will also make my setups more portable, since anchored paths work regardless of which directory I run from.
5. What I learned about Agentic AI and DevOps
My understanding of DevOps changed significantly. I used to think AI's role was answering questions or generating snippets. Now I see that Agentic AI can follow structured, multi-step workflows — planning, executing, checking results, and correcting itself. But the bigger lesson was about control: permissions and tool restrictions are not bureaucracy, they are safety engineering. Giving an agent Write access it doesn't need is the same mistake as giving a junior engineer production credentials on day one. Reusable skills reduce manual work, subagents divide responsibilities cleanly, and MCP extends reach into real systems — but human review still sits at the top. A DevOps engineer's job is shifting from "doing every task manually" to "designing the workflow, setting the guardrails, and reviewing the output."
6. My Week 2 highlight
My proudest moment was getting MCP and the agent working together for the first time. Using the GitHub MCP connection, I asked the agent to fetch a README.md directly from GitHub — and it pulled the file in real time, straight from the live repository. That moment stood out because the AI wasn't guessing or working from assumptions or training data; it got its information directly from the actual source. Watching that happen made the whole architecture make sense. It wasn't magic anymore; it was a system I had built and could reason about.
Week 2 taught me that the hard part of Agentic AI isn't the AI — it's the structure, the permissions, and the discipline around it. That's also what DevOps has always been about.
Top comments (0)