Thoughts while building Sanctum: Agents follow patterns really well. Give em good ones.

When they are missing, there is a solid chance that it will write code like I did when I was studying computer science at university, building an entire app…in a single file.

Creating well established foundations for greenfield projects has always been an avenue I’ve walked down. From my time co-facilitating a JavaScript Meetup with React / Redux boilerplates, to my time at Meetup creating full blown production ready monorepo boilerplates for greenfield Serverless projects during our AWS migration to now giving my clients (and myself) really solid launch pads for agentic development.

I’ve been using AI for development since GPT 3.5, in 2022. With the rise of agent harnesses throughout 2025 & 2026 I’ve been experimenting and thinking about a series of methods that I think are super important.

Building Sanctum, I got to really see how the rubber met the road, and wowwee, meet it, it did! This development process has been absolutely joyful.

-(Aside: Throughout this I was often sitting multiple hours in meditation a day because of the plan ♻️ dev cycle. A post for another day, and maybe a form of a single day “retreat for builders”).-

Having been a fan of monorepos for product teams in the past, I am even more so a fan of them for agents in the now. Not just for code, but also for docs and an agentic workspace for research and product decisions.

Docs are solidified artifacts that continually inform what we are building and its ethos. A full product & design spec, app store listings, etc. Docs are authoritative, current and binding.

The agent workspace is most importantly an Obsidian Vault (mention this in your agents file!), it contains research, projects - which are outputs of research, and working spaces for me to come up with decisions, copy choices, experiments, framework investigations, product comparisons -etc.

I browse and edit this space within Obsidian. This is a great way to work in my repo with the work that isn’t code.

For codebases, feature based architectures SLAY. Pre agentic dev, I would really beat the drum on domain / layered architectures, grouping code by function. With agentic development, I am finding feature based with layered flavors to be the way.

LLMs LOVE this, and most importantly, follow it. This provides the proper context for the agent at the right time. It’s token efficient and helps provide a clean, nutrient dense environment for the LLM turns.

This is honestly the tip of the iceberg but overall I am extremely pleased with how this is all turning out. I feel that Sanctum is a solid application, the codebase is nice to enter, features drop in and are iterable, edge cases are patching surgically and I feel no weight of tech debt. We’ll see how it is in a few months 😂.

I’d love to hear what others are doing - what patterns are you establishing and why?