Reality hit
I've shipped products, ran real businesses. But somewhere along the way the Rust/Go/agent-harness world was always sounds fancy to me except their concepts, and I was treating it like a side hobby — reading threads, skimming docs, feeling productive for an hour, building nothing in that direction.
Then the contract ended. No more company laptop, no more "I'll do it tomorrow." Just me, a bedroom studio, among the ocean of information and a decision of building up something cool to learn from it on the way.
So I did. Terminal became my whole world. C came back into my life. Rust and Go too — I fell back in love with both after years away. And in the middle of all that, a project called Mew just started growing on its own.
I call the way I learn as sandbox learning - Not the term I composed of course, this happens a lot out there in the Internet now.
What is sandbox learning?
A sandbox is a zone where you can try, fail, break things on purpose to learn from it.
For me, sandbox learning is:
- Pick one thing. Not ten.
- Build a tiny project around it. A real thing, not a TODO app.
- Read the manual because something broke, not before.
- Keep the box small enough to see the whole thing. No microservices, no premature abstractions.
- Feed your project everyday and see it grow up
That's it. A safety place you can play around with your craft.
That's the builder's lens — but the same idea applies anywhere
For example if you are an artist - canvas could be your sandbox; if you are a writer - substack or blog pages are; If you are a music producer/artist, send me your Soundcloud.
Information overload — what actually worked
I was learning in environments that were never designed for learning. Infinite feeds, recommended-for-you tabs, context switches every five minutes. Even with the best intentions, the wrong setup eats you alive.
So I rebuilt the whole dev environment around the way I actually want to learn — not the way the platform wanted to keep me scrolling.
Three moves did it:
1. Killed the microslop OS. Left Windows, installed Ubuntu on every laptop I own. No more "Recommended for you" feed, no more forced updates, no more "Hey, try this new Edge feature." Just the system, the core kernel.
2. Moved into the terminal. A blinking cursor and a prompt. It forces you to intend every action. No clickbait, no algorithm, no infinite scroll. Whatever you don't type doesn't exist.
3. Picked the three things that matter. Rust, Go, agent harness engineering. If it's not in those buckets, it doesn't get a tab open. Guardrail, not restriction — I still follow the same people, read the same papers, just through intentional channels, not algorithmic feeds.
What I learned to build?
"Learn the Rules Like a Pro, So You Can Break Them Like an Artist." — Picasso
Harness agent. Learn the concepts, read docs from Anthropic, Mastra, Langchain etc, conduct those main idea into the knowledge that I can use.
C and C++. I wrote a neural net in C++ as a beginner, and that's the whole point. Manual memory, no borrow checker, no GC. Just me, malloc, and a segfault at 2 AM. It's the hard way — but the hard way gave me the foundation to actually get Rust later instead of just fighting it.
Rust. I hadn't touched it in years. Picked it up again and the borrow checker made sense this time. I wrote Learning Rust in 2026 because the muscle memory came back fast once I stopped fighting the language. For now Rust agent framework is not that as mature as the other languages but nevermind, I got the concept, just build it - at the end of the day they are just binaries.
Go. Very lightweight, syntax looks lean for writing MCP with available SDK. Go for it!
Treat AI as your copilot
I use AI every day, but in an economical way. Perplexity, OpenCode, Open Weight models — they're how I move fast.
However, learning with AI is learning with a bag of bias strapped to your face. These models invent APIs that don't exist, hand you solutions that "look right" and break in production, and pattern-match your question to the most statistically common answer — which is rarely the correct one. If you don't know the fundamentals, you won't even notice you're being lied to.
Rules I won't bend on:
- Verify everything two ways. Open the real docs. Run the config block. See if it actually does what it claims.
- Know the core cold. Language fundamentals, SDLC, architecture. The boring stuff is what lets you catch the bullshit.
- Do it by hands to understand the concepts first, then enhance your productivity with AI.
Sandbox learning with AI is fine. Sandbox learning only with AI is a trap — you end up with a codebase you can't defend.
Then Mew happened
I'd been FOMO-ing about agent harnesses for months. New framework every other week, new protocol, new "this changes everything." I felt productive. I built nothing.
Then one night I just started. Rust workspace, Go MCP Server, TUI, and a name I'd been carrying for years: Mew. - Yes every developers love cats.
Mew is an AI agent harness can manipulate the ADN of other software:
Mew migrates legacy code, clones features from one product to another, and reconstructs the parts of an application that are painful to write by hand. Inspired by Mew in Pokémon — copy the original, then evolve it into something stronger.
It reads a codebase, reproduce it, understands it, transforms it, and verifies the result behaves the same way it did before - short words as reverse engineering tool. That last part is what made me build mew-skills as a separate repo — a spec kit for differential testing and regression scaffolding. Deep dive on those later on the other blogs.
Why this matters?
I started Mew with a CLI flag and a panic. Two months in it has a workspace, a TUI, an MCP layer, differential testing, and a roadmap I can actually see. Every commit is something I understand because I wrote it, broke it, and fixed it.
Pick one thing. Make it small enough to see. Build it in a sandbox where deleting everything is an option. Stay there until it does something real.
Mew is open source. If you've been wanting to learn by building the same way, I'd genuinely love to collab — a spec kit you can use today lives at tripplen23/mew-skills. Run it against your own codebase, improve the regression tests, send a PR.
I'm pushing as hard as I can to get Mew into everyone's hands. Goal: GODLIKE. We'll see.
Postscript
That's where I am now. Still learning, still breaking things. If any of this resonates — drop a star, open an issue, or just say hi. The sandbox is more fun when someone else is in it.
I would like to say thanks Sang Dang, my bro who sponsors me OpenAI credits on the journey. Thanks Mew for not having any segfaults since last week.
See you in the next one.