On this page
  • AI
  • Agents
  • Pi
  • Codex
  • VSCode

Truthfully, I’ve always been a bit skeptical of agents. Why would I want to hand off decision-making, potentially important data, filesystem access, etc to a system I don’t control or have a way to ensure alignment? What can agents do that I can’t? What is the benefit to using an agent for half of the tasks I see them deployed for? Combine that with my ever-inundated reddit feed of people complaining about token usage limits and harmful agents, and the overall negative sentiment I see towards AI on social media, and agents are honestly something I’ve just been ignoring since their inception. Sure I followed the news about them, read about MCP, TOON, Skills, Windsurf, Cursor, Warp (shoutout!), etc, but I never even attempted to use one.

So what changed? Honestly, I have 0 idea: one day I just wanted to play with an agent for fun, see what it could do. Here’s how that went:

Day 1- Qwen3.6 on my Mac

For some reason, I immediately bypassed the much-more-accessible frontier/cloud path, and went straight for a homebrew system.

“Wouldn’t it be cool if all of this was just mine? I could do it on a plane!”

I have a M3 Max Macbook Pro with 32GB of RAM, so I knew I had an OK chance at successfully running a half-decent lower-weight model, so I did some research (asked Gemini to summarize Reddit’s opinions) on what models were best suited for my needs: learning Rust. It came back to me with Qwen3.6.

Building a Crawler…not the cool kind

After doing the math and learning a ton about what dictates RAM requirements, I was naively optimistic I would end up with a semi-productive system using Qwen3.6-27B-Q4, or a 27 Billion parameter Mixture of Experts (MoE) model with 4-bit quantization. All of those words I just spit out were completely new to me and I had to do quite a bit of reading; my Medium subscription finally paid off! Anyways, if you’re at all familiar with local LLMs or hardware requirements in general, you can see where this is going: 27 is very close to 32… I ended up with a blazing 6.4 tok/s at the model’s peak performance, yikes! It was completely unusable, so I quickly pivoted towards a smaller model (which Gemini recommended from the get go I might add).

14B saves the day!

I went with 14B, still with 4-bit quantization. It ended up being 4-5x faster and somewhat usable, still slow, but my thoughts at the time were:

It’s just so cool that this is running on my laptop, I kind of don’t care that it takes forever.

And then I tried hooking up a harness to it… This is the part of this post that I’ve debated writing about the most, and I’ve decided to keep it brief. I went with Aider initially, said hello, and then it started recursively searching my entire filesystem, didn’t play nice with git, etc. I don’t blame Aider there though, more the model/user error. I wanted VSCode integration, so I looked at Continue and Cline, quickly gave up on both. Then I saw a talk on Pi… and I was hooked on the modular approach, minimalist approach (more on my thoughts on barebones harnesses some other time), and “If you want Pi to do something new, just ask Pi to build it for you”. How cool is that!! That’s when I quickly realized that Qwen just wasn’t going to qut it. I asked it to scaffold out a basic Rust project for me, and not only did it spit out some really weird/invalid filenames, it straight up hallucinated a Rust version in the .toml that doesn’t even exist…Not great for my pair programmer! At this point I had two options: buy some really expensive hardware like a DGX Spark, or a much cheaper subscription. My heart wanted option one, but let’s be real here, I went with option two.

Day 2-30 Codexmaxxing

I chose OpenAI. This was right at the time that 5.6 came out, and I just prefer their current pricing model/usage limits, I didn’t want to be limited or afraid to ask my agent to do something due to cost. I could write a whole post on this, but it would contain nothing that hasn’t been said by people far more knowledgeable than me. I’ve primarily been using the ChatGPT Mac app and VSCode extension, but for quick tasks I use the CLI. I’ve also been continuing to use Pi, more on that later. Honestly: my experience has been fantastic, eye-opening, game-changing, etc. I now listen to Agent-focused podcasts on my commute and in the gym, and I come home from work genuinely excited to build, even though I already stared at VSCode for 8 hours that day. Great, end of blog post, see you guys next time!

Kidding, here are some of the things I’ve learned and some of my opinions on the current state of things.

Learnings.md

I’ll split this up into quick bullets, and some longer subsections.

Rapidfire Realizations

  • LOT better at coding than 2-3 years ago (to the author: duhhh)
    • Shoutout Maizey from EECS485
  • Can finally write and understand Svelte 5
    • No more pasting llms.txt into every context window ever
  • Really dumb at filesystem searching
    • “Allow ChatGPT to access your iTunes library?”
  • Slow at writing git commands
  • AI READMEs are straight slop
  • CSS is genuinely almost not even worth it
    • Probably because all of the training data is also bad CSS
  • Obsidian is pretty great for Agent organization
  • I’m awful at using Pi

The 1-on-1 Programming Tutor you’ve always wanted

Title, I’m learning Rust through building codegraide and Codex is great at outlining steps, going one small chunk at a time, not accidentally writing the code for me, etc. I’m using this term very generously here but, quite frankly, it’s like having an expert sitting next to me that I can ask any question, no matter how dumb. I will say however, it tends to over-engineer a LOT and tends to hallucinate extreme hypotheticals that don’t matter/will never happen.

5-year long Notebook Idea Backlog -> MVPs in a few hours

I’ve been able to turn all of my half-baked and full-baked ideas into (slightly) functional prototypes, or even just research collections for further reading. This is pretty cool, but also dangerous, if you thought my 50 unfinished projects or whatever was bad before, it’s about to get so much worse. Oh and by the way, they’re ALL going to be desktop apps built with Tauri. Kidding on the last point, but for quick drafts/ideas where I genuinely don’t really care about quality/slop, it’s nice handing the coding reigns to someone else and just making design decisions.

Planning is…mid

/plan, from my usage, has been a little annoying to use at times: the constant user-requested input blocking (just gets tiring), some of the recommendations it comes up with, Sol-High’s LOVE of over-engineering an MVP, and the actual usefulness of the plan itself. I tend to question how useful it actually is; curious to see how it evolves or slowly deprecates as pre/post-train cycles only get better and better at supporting agentic workflows.

Memory might be overkill, but I love it

For each project, I have an obsidian vault with an Agents section, and a user section. I store everything here, or rather tell the agent to. Design decisions, change logs, skills, you name it. I’m aware of the decreasing need for custom boilerplate prompts/skills/etc, but from an observability perspective: it’s amazing for me. I can go in there and correct things, read the history of the project, and so much more. And it’s all in a pretty graph in Obsidian. Obviously I’m just scratching the surface here, lot’s of research to be done to get better/more efficient at this. I’m also curious to see what plays out in the agentic-memory startup space, so many different possibilities!

It might be time to say goodbye to VSCode

There’s a very real chance I switch to Zed or something else in the coming weeks. VSCode just doesn’t feel agent-native, if that makes any sense. I’m not using, and don’t want to use, Copilot, so my Codex integration mostly only comes through the chat window on the side, I feel like I’m missing a lot of features. Copy and pasting terminal output into the chat window gets annoying. I’m sure there are solutions to this, maybe even extensions that I’m just not aware of, but right now VSCode is one of the biggest rubs in my day-to-day.

Three Things I built

  1. Reteyre
    1. Local dashboard for aggregating all of my financial providers (banks, stocks, etc). Tauri, Svelte, Shadcn. Didn’t want to trust a 3rd party source with my financial fingerprint.
    2. Works OK, lots to improve on.
  2. This site.
    1. Gave Codex the basic design, said I care only about speed/bundle size and having a fast site. Codex suggested Astro, and I let it take it from there, with lots of input along the way.
      1. I needed a new site, and didn’t want to build one. I honestly don’t really care about frontend programming and not having to build a decent site by hand was my main motivating factor behind paying for the subscription.
    2. I generate my content with Markdown files so writing is as frictionless as possible, I just do pnpm create:new and interact with a CLI that scaffolds whatever content I’m adding to the site.
    3. Had to do a lot of manual style-wrangling, Codex’s attempt at grids and CSS got messy real quick.
  3. Codegraide
    1. Hand-written as much as possible, but I don’t have a ton of time on my hands.
    2. Number one thing I’ve noticed:
      1. It’s meant to be a fully modular design w.r.t. languages, each language gets it’s own analyzer, but Codex loves to bleed language-specific terminology or features into the core code
    3. Codex is pretty bad at naming things: usually verbose, obscure word choices, just not very clear at times

Closing Time

Thoughts on the Future

  • Would love to see Code Review improve w.r.t. code the model itself generated
    • Hopefully codegraide helps someday…I wish.
  • Claude Design is far better than anything OpenAI currently offers, will this become a priority?
  • Still not a fan for using AI for deterministic/scriptable tasks.
    • Use AI to write the script instead!
  • Everytime I do anything package-manager related, it seems like I have 20 packages out of date!
  • Can we all agree to have one internet crawler and then we just pull the data from there?
    • My projects and sites should NOT be getting the amount of views they are- all bot traffic
  • I forgot to talk about my Pi experience at all…

Hope this post was at least somewhat insightful/funny, lots more to come!

While I frequently utilize AI for a variety of tasks, including maintaining this site, all of these words are my own and I did not use AI to generate this post.