how I'm using agents (pt. 1)
I read Paul's post about how he uses LLMs last night, and I've previously read Dylan's post about spinning plates.
After reading Paul's post yesterday, I was thinking it might be worth my sharing how I'm using LLMs / doing 'agentic engineering' too. I feel like I'm pretty in the trenches with it and have been since ~October 2025. Where I work, I consider myself one of the engineers trying to push the boundaries of how we use it.
I've basically not been writing code by hand since November. Occasionally, I try to do some just as practising craft or fixes that I can do even quicker than an agent (one-liner YAML sort of changes), but otherwise I'm all in and have been for a while.
I'm still reviewing the code, although I'm quite hand-wavey about the implementation. I'm doing the system design up front, providing context and use cases, and outlining how we're going to validate it, which makes me pretty confident about the output.
I then manually test the changes pretty frequently. I'll either run locally (where possible, we're not well set up for it) or deploy to our dev environment to poke around. I've got some deterministic test harnesses that help a lot, too. Especially great for the event-driven system I own.
I'm frequently on the lookout for new tooling, skills, or harnesses to try out, to help me maximise.
I'm using Pi as a harness, and it's by far the most effective I've used with minimal effort. Most tooling works great with it. I pair it with RTK to get the most out of my tokens, and haven't had any downsides.
I'm currently using GPT 5.6 Sol as my daily driver and Fable 5 as my planner for bigger things, but then offloading to GPT again. We're given very, very generous monthly budgets, so I tend to just keep things on xhigh/max effort. I do end up getting through almost all of my Claude budget, and it feels pretty wasteful, but I'd rather use tokens in getting it right the first time than waste them going around in circles, and I'm pretty sure that is the best way to do it.
I've got more to share here, but need to stop for the day! Will make a follow-up on more specific workflows and what works best for me.