Hey there! I'm late on these because I didn't go to a coffee shop yesterday morning, and now I'm fully discombobulated. Oh well; things happens.
microclimate progress!
I'm proud to report that I manage to work on a project for a two weeks in a row! And even more impressive than that: It works! I've successfully implemented chatting with other folks via a LiveKit instance.
Now, I haven't gotten it to work with my self-hosted instance yet, but I'll figure that out in time. My #1 goal was to prove that it could work, and it does--with audio quality fairly similar in quality to what we had at our Tier 3 server boost level. Volume and background cancellation leave something to be desired, but we can solve those problems later.
There aren't any spectacular technical challenges that we've had to overcome at this time, so I wanted to spend a little bit talking about how I've been exploring agentic coding.
Glorified search
Historically, I've been more of a breadth-of-knowledge (vs depth-of-knowledge) person, so I've really struggled when confronted with a challenge that involves compounding factors that I'm unfamiliar with. I've been prone to bail on libraries, runtimes, and even languages when I get too stuck. Friction has resulted in many a dead project.
I've been pleasantly surprised that the agent-as-search use case has helped me persist with this project at multiple points. Programming has always felt like a means-to-an-ends to me, and I've never particularly enjoyed the in-between. Maybe LLMs can help me change that?
Damn, glorified search got hands
Up until this particular project, I was leaning towards never delegating code writing or debugging to an LLM. I hadn't really... vibed with vibe-coding. It felt like I was robbing myself of an opportunity, and I was frequently unhappy with the code they produced.
I've had a small change of heart with that. Sometimes it's necessary for me to delegate writing and debugging, when I'm too frustrated, or when there's too much uncertainty.
Making lunch or taking a walk can help with bridging the gaps between moments of flow. Unfortunately, I have an avoidant personality--if there's a challenge that I'm struggling with on the other side of my break, I might just never come back to address it. Many of my projects have died this way.
In that situation, delegating tasks to an agent has helped tremendously with keeping my focus on microclimate. Sonnet 4.6 has helped me debug issues and scaffold out modules where I fear previously I would have lost interest. It can read so many logs and get insights so much faster than I can--for reference, I struggle greatly just to follow the steps in a recipe.
I'm still figuring out how much I feel like this is a crutch. I don't get the same gratification out of an LLM-delegated task than I do when I ask an LLM how I should solve a problem. That said, microclimate is the first project where I believe that agentic coding has been a net benefit to me.
Okay, I thought these would be quick, but I've probably spent 3 hours writing and rewriting this log. I need to be done now.
If you got this far, I hope some of that made sense. Thank you for reading, and I hope you have a great week!