I Didn't Get the AI Agent Hype. Then I Built a Second Brain That Knows Me.
I was skeptical about AI agents. Six months later, my agent knows my writing voice, my job search, my learning goals — and I'm getting more done than ever.
I was wrong about AI agents.
Not “slightly miscalibrated” wrong. Not “had the right idea but bad timing” wrong. I was the kind of wrong where you roll your eyes at the hype and go back to your terminal.
The pitch always sounded like science fiction. “Autonomous agents that can write code, manage your files, learn your preferences, and operate across your entire system.” Sure. And my Roomba is going to organize my closet.
Then I installed Hermes Agent on a Hetzner VPS six months ago. Not because I believed. Because I was curious enough to be proven right about my skepticism.
I was not proven right.
The backup brain I didn’t know I needed
Here’s what actually happened: I started using it for small things. Explain a Rust concept. Generate a quick script. Nothing I couldn’t do myself, just faster.
Then it started remembering.
Not in a creepy way. In a practical way. I told it once that I prefer short, direct answers over polite fluff. It stopped doing the fluff. I mentioned I was job hunting and targeting remote Python roles. It started framing technical explanations around interview-relevant concepts without me asking. I saved a convention for how I format my resume files. It never got it wrong again.
That’s the part the demos don’t capture. The agent isn’t just responding to prompts. It’s accumulating context. Every correction, every preference, every “actually, do it this way instead” — it persists. After a few weeks, it felt less like a tool and more like a collaborator who’d been on the project since day one.
Last week I asked it to review a job description and suggest how to frame my Outreachy experience. It pulled up my contribution history — 110 commits, FastAPI, HTMX, OpenShift — without me mentioning any of it. It knew because I’d told it once, months ago.
I sat back and thought: okay. I get it now.
What I actually use it for
The sales pitch for agents is usually about coding: autonomous PRs, multi-file refactors, building entire features from a prompt. That stuff works. But it’s not why I keep coming back.
Learning things I’d otherwise avoid
I’ve been meaning to learn Rust and Go for two years. The problem with self-directed learning is that nobody tells you when you’re building bad habits. My agent doesn’t replace a course, but it’s an always-available sounding board. I ask it to explain a concept, show me idiomatic patterns, review my code. The fact that it knows my Python background means it draws comparisons I actually understand.
Yesterday I wrote a small Go HTTP server and asked for a review. It pointed out three patterns that aren’t wrong but aren’t Go-idiomatic — things a linter wouldn’t catch but an experienced Go dev would. I wouldn’t have learned that from a tutorial.
Crafting job applications that don’t sound generic
Job applications are a grind. Tailoring a resume and cover letter to each role takes time, and after the fifth one, you start sounding like a template.
My agent has my full work history, my writing voice, my career goals. When I give it a job description, it can pull out the relevant experience, rephrase bullets to match the role’s language, and write a cover letter that sounds like me — not like someone who discovered thesaurus.com.
I still review everything. I still make edits. But starting from an 80% draft is radically different from starting from a blank page.
Operational stuff I’d normally procrastinate
System updates. Log rotation. Checking if backups actually ran. The kind of maintenance that’s important but never urgent. My agent runs on cron and handles things I’d forget about for weeks. It’s not glamorous. It’s the difference between a server that degrades and one that stays healthy.
The part that surprised me most
It’s not the capabilities. The capabilities are impressive but expected — LLMs are good at generating text, we know this.
The surprise is the accumulation. Every interaction makes the next one better. The agent learns my tools, my file structure, my naming conventions, my preferred package managers, my writing voice, my career goals. It’s not like a smarter Siri that resets between conversations. It’s like a junior engineer who joined six months ago and actually paid attention.
I didn’t realize how much mental overhead I was spending on context-switching until I didn’t have to do it anymore. “Here’s what we’re working on, here’s what I need” — and the agent already knows the “here’s what we’re working on” part.
The honest trade-offs
This isn’t a magic wand.
It gets things wrong. Sometimes confidently wrong. You need enough domain knowledge to catch the mistakes, which means the tool is most useful in areas where you already have some competence. If you’re a complete beginner in something, the agent can mislead you without you knowing.
It costs money. API calls aren’t expensive per-token, but they add up. I’m spending maybe $15-20/month on API credits. Worth it for me. Might not be for everyone.
There’s a learning curve to setting it up properly. Getting the config right, understanding what tools to enable, figuring out the memory system — it takes a weekend of tinkering. The CLI experience is powerful but not polished. If you want plug-and-play, this isn’t it yet.
And honestly? Sometimes I wonder if I’m offloading too much thinking. When the agent remembers everything, do I remember less? I don’t have an answer to that. I just notice it.
The series I’m starting
This is the first post in what I’m calling AI Agents for Daily Life — not the sci-fi version, not the VC pitch, but the actual lived experience of using an agent as a thinking partner.
Next posts:
- How I connected Hermes Desktop (native GUI) to my remote agent via Tailscale — zero open ports
- Using an AI agent as a career coach: resume tailoring, interview prep, skills roadmapping
- Automating the boring server maintenance you’re definitely ignoring
I’m not selling anything. I’m just a software engineer in Kenya who was wrong about something and wants to document what I’m learning.
If you’re on the fence about agents: the demo videos are hype. The real value is quieter. It’s in the small moments where you realize the thing you’re talking to actually knows what you’re working on.
Follow the series at nyagah.me. Find me on X/Twitter or LinkedIn.