Engineer at Grw AI. Production agents for sales, side projects in finance. Bike person, bad guitarist. Half-formed thoughts on what's shipping
Wellington, New Zealand
Joined January 2008
- Tweets12K
- Following1.1K
- Followers1.1K
- Likes5.9K
I don’t think enough about how wildly software Eng has changed in the last ~2 years
Maybe I’ll sprinkle some more console.logs in my next pr for old times’ sake
strong agree
• comments go stale very quickly and confuse agents and people
• they are often used by agents to explain workarounds that would've been better off actually fixing the code instead
• they are not enforceable and waste your agents' tokens
my point is that moving prose (comments, rules, agent.md, skills) into deterministic lint rules, better API design/types, is far more effective, enforceable, and token efficient.
comments are simply a crutch that agents will happily misuse
Here's some takes on Agentic Dev:
- Force your agent to comment like they're being graded. Allow them to clean up nearby comments too
- Ask for the subtractive fix
- Delete tests that restate the code (Tautological tests considered harmful)
- use t3code nightly
what's yours?
Other side of this is the most valuable devs from 2+ years ago might not be the most valuable today with how far software dev has shifted
⬅ Mathew Hartley retweeted
Just told Claude to have some "me time" and work on something that it would enjoy.
Got inspired by UniiChat by @VictorTaelin so vibed a plugin for my Hermes bot - github.com/fourseven/hermes-…
(still back porting my history across though 😅)
Lemme know if this is useful for you
⬅ Mathew Hartley retweeted
🗞️ the world hasn’t figured out yet that you can literally just fix everything with a Nix overlay
In this blog post, I go into how I use Nix, why you should use it, and in the tweet below is a runnable demo showing all of the concepts in this post ghuntley.com/nix/
⬅ Mathew Hartley retweeted
It's actually INSANE how prescient this was.
They made this in 2011! This is basically what it's like to use computers now. It's barely exaggerated
⬅ Mathew Hartley retweeted
So, the work that @dhh has done with the campfire shootout is interesting in a few ways:
1. He's shown how very easy it is to port small-medium legacy webapps to a new language using AI with minimal involvement. (I'd done something similar this year handing Claude the source code for system shock 1 and porting the renderer to JS.)
2. He's shown that these ports will suffer in ways predictable to experienced users of those languages but hard to catch otherwise.
3. He's shown (really, reaffirmed) the easiest way to get information on the internet is to loudly say something wrong and bait domain experts into replying.
4. He's shown that even unpalatable languages (e.g., Rust) with nice performance benefits can now be attempted in places where nobody sane previously would've done so.
These are all useful results!
Agents: The Good Parts
yeah i feel like there are a lot of parallels between systems that provide constraints (eg type systems) and using agents effectively and at scale. when you are dealing with extreme volume that you cannot control directly you have no choice but to shape the environment. i was inspired by my own journey trying to use typescript effectively
no.lol/2019-12-27-type-narro…
Honestly not surprised - there’s a thought on how (generously) the same prompting produced different levels of “good” between languages but yeah maybe vibes more than science here - I’m keen to see experts open PRs to make this representative
Come on. Your agents run all SQL queries through a single genserver, with all queries awaited via `GenServer.call`. This is a moronic Elixir implementation. I've staid fairly agnostic on the whole "rewrite everything in rust" thing but if this is the Elixir that your agents are writing, I'd hate to be a rust expert looking at their rust.
github.com/basecamp/once-cam…
this 1000x, cool teams doing cool stuff makes me forget I'm at a job
if you're job searching right now: choose the team, mission and approach that resonates most, and where you'll have the most fun.
not just the buzziest or apparent leader. with model advancements, products reinvent themselves every 3-6 months, so the leaderboard resets often.
if you're chasing hype, you'll be unhappy once you're (inevitably) out of the hype cycle.
⬅ Mathew Hartley retweeted
Just at the time everyone else is getting rid of this feature. Goodness Google, what happened to you?
Antigravity 2.0 now features a dedicated planning mode, just like the Antigravity CLI.
Type /plan and the agent steps back to think through the task, and conduct extensive research, before generating an implementation plan for your review. The agent will ask for your approval before it begins executing on the task.
You can also ask for a plan naturally in your prompt, for a lighter version of a plan.
Put down my thoughts on the Flash revolution: DeepSeek's v4 Flash 0731 checkpoint got me running a general purpose (Hermes) agent for about a dollar a day, which is a long way from the $200/month OpenClaw era.
mathewhartley.com/blog/2026-…
