A Month Inside the Agent Loop

The last 30 days involved a lot of tinkering. A non-exhaustive list of what the month taught me:

• Pi pulled me into the guts of the agent loop. I understand harnesses, extension points, and where the real leverage sits far better now.
• Rewiring Pi extensions finally forced me to learn enough TypeScript to stop dodging it.
• Raw terminal harnesses teach faster. They give up comfort, but they give me control.
• The prompt-to-project leap can happen violently fast when the loop stays close to the metal.
• oh-pi made Pi practical for daily use. Better visibility into branches, active models, and context-window burn changes how I work.
• Visibility matters. I want to see what model is running, how much context I am burning, and what the bill is about to become.
• Model routing is a first-class problem. I wish I do not have to blindly trust codex or gemini to pick the right model, review cycle, or spend profile for me.
• Flash models are good scaffolding engines. Heavy models earn their keep when the logic knots up. I’d go one step more and say that gemini-3-flash is the best scaffolding model.
• Swarming everything still feels like overkill, even if the larger issue -> design -> spec -> plan -> tests -> implement -> review -> test -> push loop still feels right.
• The open question is cadence: when do I actually want more passes through test, implement, and review, and when am I just indulging ritual?
• Writing skills in English is liberating, but I still hit the limits fast when I need deterministic behavior.
• ant-colony is a great name for a sub-agent execution mode.
• I like digging into details, sometimes more than shipping, because understanding the system changes how I use it.
• OpenAI and Google models still beat local models for most serious coding-agent work.
• Local models may get there. They are not there for me yet.
• I burned through $20 plans much faster than usual. The learning curve and loose routing both made the meter run hot.
• I still cannot justify the $200 plans. Cost control is part of the product now.
• Gemini CLI got materially worse for non-enterprise use. I get why Google clamped down, but the downgrade still stings.
• Going under the hood made me more articulate about what I want from these tools. Friction is still one of the better teachers I have.
• There might be a future world in which LLMS and agent harnesses will take care of all coding tasks. However, for now, they can be tripped up by dumb stuff, including just stopping execution by accidentally pressing the Esc and resuming.
• However, they are particularly good at generating a LOT of code, very fast. This is particularly useful when most of the code required for many things is scaffolding.
• This is why at least for development, LLMs actually make the process more interesting as it allows one to put energy into the interesting parts of development.
• That said, life is never that simple. I wonder if the interesting parts of software development are interesting because they are relatively more interesting than the mundane parts of it?
• Trust me when I say that you need to be able to review the AI generated code. There are sometimes egregious mistakes that can be caught even by a casual review. I’ve now stopped auto-merging any LLM generated code in my repos.
• It’s weird, but the more you use LLMs, the more they make you skeptical. They are at once incredibly useful and dumb. One thing they are : confident!
• I don’t know if AI will take over software development or any other job, really. However, for anyone who likes their craft, it is an augment.