Chris Armstrong

Arrggh trying to get opencode to read skills from a custom path

Having a harness like opencode which gives full visibility into subagents and their full prompts is essential to being able to optimise things, and having control keys you strip out the useless and wasteful bits. I’m definitely a fan of minimal frameworks and libraries for the software I’m building.

The thing with this approach is you can create really complex tooling that adds little. I’m creating tools from scratch to make sure I really understand how it works and improves things, so you can strip it right back to just what you need for maximum efficiency (wasting tokens just slows you down).

I’m really keen to find an affordable host + chat interface to run multiple agents, if not in teams, at least driving small plans myself. Zocomputer looks interesting, but finding the time 😭

I’m actually finding the most success with flexible harnesses like opencode and validating things with cheaper models (where I have tokens to burn rather than exhausting my quota in a Claude based Ralph loop after one phase of a multistage plan). Opencode’s Zen Go is a huge boon for this.

At the moment, spending much of my time at 4 (context engineering) and 5 (tools/skills/mcp) quite a bit, with a lot of ongoing challenges describing tools in a way that I can get the LLM to use them.

The 8 Levels of Agentic Engineering — Bassim Eledath AI's coding ability is outpacing our ability to wield it effectively. That gap closes in levels — 8 of them. Here's the progression from tab complete to autonomous agent teams.

I read an article like this, and it really resonates. I’m picking up bits from levels 1-7, but still finding myself trying to master 4-7 at the same time.

www.bassimeledath.com/blog/levels-...

I’m also hoping to have some good news around ocgtk relatively soon too.

I have been a bit quiet on the ocaml front, but have been working on a private project using miou vif for a website/cms/newsletter/events management thing (are there any robur coop peeps on here?) thats quietly coming along

I'm based in Sydney, available on-site/hybrid or fully remote. Open to permanent and contract positions.

I'm open to new opportunities at the moment. Founding engineer at a fast-growing finance-based startup and working in startups/scale-ups over the past 10 years.

Specialties are DevOps, cloud architecture & observability in product-driven environments.

Node/Typescript/AWS/React, also OCaml.

A screenshot of a Claude CLI session that reads: * Sauteed for 51s > for the love of god do not run it Noted. I'll only edit the file, no execution.

me using Claude to investigate a nasty memory leak

AI Didn't Break the Senior Engineer Pipeline. It Showed That One Never Existed. Most organizations never had a model for developing engineers. They had an environment that produced growth by accident. AI just made the luck run out.

If coding is your favorite part of software engineering, keep coding. AI doesn't stop you. But if coding was the only part you were good at, that's a different conversation. The job was always bigger than the code. #bransoncognac blog.bryanl.dev/posts/ai-sen...

Yeah I need to try codex a bit more for code review, maybe use it with a GLM or Sonnet executor

Detailed to a degree unlike other models, not comprehensive like every model

(Why I strongly believe in using cross model review at the task and phase level)

I apparently have Tesco Pitas down as well.

Made bagels today 🥯 🤤

# isn’t perfection (lesser models and Claude generate imperfect code under all sorts of constraints and prompting in my experience)

I’ve built my own set up for agentic coding using Claude but I would say it’s perfection; I’m constantly refining it.

If I just use a minimal harness and get Claude to build it, I need to refactor it anyway because it’s an unreliable mess

Soon to be MIT/Linux, because they’ll be vibe porting the GNU toolchain to Rust and it won’t be having a FSF-prescribed license

Mature companies I talk to have significant plans to move off any stack that isn’t otherwise JS/Python/Java/.NET (PHP or Ruby especially). Performance sensitive stuff goes to Rust, but only if necessary.

“The market” and AI have definitely converged the language and library space for new startups to Node.js and React in Australia. Everyone is scared to build outside of them because the talent pool here is so small and its hard enough to hire.

Discovered only through conversation at a meetup last night that a mutual friend was an Elixir developer in a previous life - apparently there was a small number of Elixir shops in Australia some time back.

In think it uses the lazy plugin manager under the hood, but gives you a structure for adding plugins one by one

GitHub - nvim-lua/kickstart.nvim: A launch point for your personal nvim configuration A launch point for your personal nvim configuration - nvim-lua/kickstart.nvim

If you’re initialising a neovim from scratch I highly recommend github.com/nvim-lua/kic...

Overproofing the sourdough nowadays. My starter is a mutant beast but it only like the good stuff

My slides (and my video, soon to be linked) from last night are here: www.chrisarmstrong.dev/posts/sydney...

Sydney AI Engineering Meetup talk - A field report on using LLMs with a little-known language Last night I presented a talk at the Sydney AI Engineering Meetup on using agentic coding loops with lesser known languages and less-capable models, of which you can find the videa and slides linked

My slides from last night (and soon the video) are linked here: www.chrisarmstrong.dev/posts/sydney...

I really do need to link to my coding guidelines though and somehow store them separately to the repository, including the dependent agent configuration

This was recorded too 🎥 so hopefully i can post a link