It is 2 a.m. My laptop is still on.
I am a final-year student working on my thesis while juggling an internship, quietly typing code between deadlines that keep stacking up. The rumor is everywhere: “Agent AI will do it all for you.” The rented room feels huge, the fan hums steadily, and a cold cup of coffee sits on the desk. The air is tight, like a guitar string about to snap.
My heart is pounding. How am I supposed to survive the last semester when everyone keeps saying AI is about to replace me?
1. The Bottleneck and the Blind Spot
Students and interns tend to fall into the same trap. They believe ChatGPT or Claude will write clean code for them, so they hand everything over instead of understanding the foundation. Some only learn theory on paper — read books, attend lectures, call it a day — then panic the moment they hit a real error. Some copy-paste code from the internet, tweak one or two lines, submit, and never bother to debug or figure out how it actually runs. All three patterns are short-lived work, because AI is not magic. It is just a tool.
Even the big players don’t treat AI agents as the engine that runs the show. GitLab 18.5 introduced the Security Analyst Agent to auto-classify vulnerabilities and the Planner Agent to manage backlogs, but both still need engineers to build the scaffolding around them — the configuration, the approval flow, the project context. Agent AI is powerful, but it is not infinitely flexible, and it still slips up like any human. What matters is how you use it, not what percentage of your job it can take.
2. Hands-On Experiment
I decided to get my hands dirty and try it for myself. The environment: Windows 11 (build 19045) with WSL2 Ubuntu 22.04, 4GB RAM, Python 3.11, VSCode 1.89.2, Docker 23.0. The CLI is the main weapon — no shiny GUI.
Three open-source agent CLIs I actually ran, not made-up stuff:
Aider — terminal pair-programming that auto-commits each change to Git:
python -m pip install aider-install
aider-install
cd duong-dan/den/du-an
aider --model sonnet --api-key anthropic=<your-key>OpenAI Codex CLI — OpenAI’s open-source terminal agent, written in Rust, running inside a network-isolated sandbox:
npm install -g @openai/codex
codex auth login
codex --auto-edit "Explain this codebase and suggest improvements"Gemini CLI — Google’s open-source agent running a ReAct loop (reason + act) with built-in tools and MCP server support:
npx @google/gemini-cliAll three actually executed real tasks, but none of them ran smoothly on their own without fundamentals. To make agents stable, I still had to write my own small utilities — shell scripts, log parsers — for the in-house systems that don’t expose a proper API, which is extremely common in the Vietnamese market. When the agent runs commands on real infrastructure, errors still show up: missing permissions, missing libraries, compile failures. That is when I had to fix things myself. No one else was going to do it for me.
In short, agents don’t take the pain for us. Even GitLab Duo Agent Platform only runs on-premise, inside the company’s own secured infrastructure. Agents are only powerful when the operator truly understands the system they’re being run on.
3. Takeaway
Optimists call the agent an assistant. Pessimists call it a threat. The truth sits in between. The only way to know is to try it yourself: take a small project, hook Aider or Codex CLI in as a reviewer in your CI/CD pipeline, let it fail, then debug it yourself — you only learn fast when it breaks. Every time AI gives you a bad suggestion, write it down. It is a hint engine. The final check is still your hands.
If you are still hesitating, just install one of the three agents above, point it at your repo, read the original docs, and run the real commands. The fear of being replaced fades a little every time you fix an error on your own.
Sources: official documentation from OpenAI (github.com/openai/codex), Google (github.com/google-gemini/gemini-cli), Aider (aider.chat), and GitLab (about.gitlab.com), combined with the author’s hands-on experience.
