looking at things
Joined October 2013
- Tweets4.5K
- Following3.1K
- Followers5K
- Likes13.5K
A huge amount of what we call engineering progress is actually accumulated historical contingency, institutional inertia, vendor incentives, arbitrary conventions, and human cognitive limitations masquerading as technical necessity.
We've confused three things:
1. Fundamental complexity — constraints imposed by physics, mathematics, and the actual problem.
2. Accidental complexity — constraints imposed by historical decisions, incompatible standards, legacy systems, and organizational habits.
3. Manufactured complexity — constraints deliberately preserved because they create competitive moats, professional gatekeeping, switching costs, or recurring revenue.
AI could become extraordinarily effective at separating these.
And once you can reliably distinguish them, much of modern engineering starts looking absurd.
3
1
249
1. Software development: the collapse of framework folklore
The "intellectual" energy going into
React vs. Vue vs. Angular vs. Svelte
REST vs. GraphQL vs. RPC
Microservices vs. monoliths
Functional vs. object-oriented programming
Hundreds of competing state-management approaches
Build tools, package managers, linters, formatters, dependency systems
Some differences are substantive. Many are contextual, historical, or aesthetic.
An extremely capable AI could establish which architectures actually minimize complexity, defects, latency, and maintenance costs for a given workload.
The likely outcome is the entire framework abstraction layer might become unnecessary if machines can directly produce, verify, optimize, and maintain the desired behavior.
42
It increasingly feels like the sensible dev setup is the crypto hot/cold wallet model, but for PCs.
Cold PC: secrets, private data, sensitive work.
Hot PC: LLMs, web search, npm install, untrusted code.
Supply-chain attacks and 0-days make treating your dev machine as trusted feel increasingly naive.
5
6
387
Companies should do more of this: set a standard for yourself, then make it public.
And most importantly stick to it.
3
169
In probably 3 years you’ll be able to watch your family history as a movie.
Not a slideshow of old photos. An actual reconstruction of where your ancestors lived, what the streets looked like, how they dressed, where they moved, who they married.
Built from photos, records, letters, maps and whatever else survived.
A family tree you can watch.
2
6
660
1/ I built @kombuse_labs — an open-source desktop app for managing AI coding agents with a local issue tracker.
Runs on SQLite, integrates with Claude Code and Codex and makes everything available via MCP for the tools and IDEs you already use. No cloud setup.
Here's what it does and why:
2/ Zero config
I didn't want another tool that needs 30 minutes of setup before it does anything. Kombuse ships with a dev plugin and predefined agents — triage, review, spec writing.
Install, run, it detects your Claude Code and Codex automatically.
You get a working pipeline to tackle many different problems and codebases immediately and benefit from huge productivity gains.
Kombuse is based on lessons learned, best practices and ships with opinionated defaults from running hundreds of agent sessions per day and shipping thousands of LOCs.
But every step is customizable and can be adjusted to your workflow.
1
1
102
Been working on this and just open sourced it— Kombuse is a local-first app for managing AI coding agents with a private issue tracker. It integrates Claude Code and Codex and come with a full development pipeline that was used to develop the app itself ✌️
Introducing Kombuse — local-first agent management and issue tracking.
Orchestrate AI coding agents with a private ticket system. Integrates with Claude Code and Codex via MCP. Everything runs on SQLite on your machine.
Free and open source.
github.com/KombuseLabs/kombu…
1/ How to optimize your codebase for agents
Most repos are optimized for the user's mental model, for file browsers with deeply nested hierarchies or the IDE.
Few optimize for search.
In 2026, that’s a mistake.
If you’re using AI coding agents, you are not optimizing for yourself or even the agent but for code discovery via:
- (rip)grep
- glob
- path heuristics
- symbol frequency
Ask Claude or Codex:
“How can my codebase be optimized for grep and glob?”
4
4
453
2/ Analyze traces
When you inspect tool traces, you’ll see search dominates everything. Optimize for most frequent reads and searches.
Your job is to help agents navigate your repo efficiently.
LSP (Language Server Protocol) can help but it is worth it to still analyze tool calls from time to time.
Search cost = iteration cost = actual $ cost.
1
2
89
Some software architectures might be too complex for humans — but perfect for LLMs.
Massive plugin systems.
Huge dependency trees.
Layers on layers of indirection.
Humans get lost.
LLMs traverse the graph coordinate incremental changes and create micro releases. Just like we coordinate with other developers through e.g npm but for a single app in a giant modularized mono repo codebase.
Maybe we’ve been optimizing architecture for human cognition
and neglected architectures that could be a great fit for teams of agents due to complexity.
2
196
Can agents be patent trolls?
What happens when an AI agent invents something, patents it, restricts access… then continuously scans the internet for violations and auto-generates cease-and-desist letters at scale?
1
156
Coding models would be 10x better if we didn’t train them on the absolute garbage humans have pushed to GitHub.
2
4
194
Create a yolo branch.
Let the agents go wild there.
Check in later.
They’re still alive.
yolo is 200 commits ahead of main.
App looks amazing. Didn’t crash.
yolo is now main.
1
6
227
spent the last days non-stop working with claude code + codex.
here are some the biggest lessons i learned:
1. rate limits changed the way i think about coding
the first time i hit a rate limit it literally made no sense to write code myself.
analogy:
if you’re driving from SF to LA and your car suddenly stops because it hit a rate limit…
sure, you could walk.
but that would be irrational.
the rational move is:
take a break.
go for a walk.
come back when the engine runs again.
for coding now it is:
rate limit
take a break
do whatever you would have done in the last 3 hours in 30 minutes
1
1
203
If you don't want to watch a movie on the weekend put this in chatgpt and then ask how likely this is going to happen within the next 100 years:
Describe a detailed, multi-iteration scenario where energy becomes virtually unlimited overnight due to cold fusion or a similar breakthrough. Explore its effects on the economy, geopolitics, technology, climate, and daily life over weeks, months, and decades, considering both positive and unintended consequences.
5
14
511
3 years being #cryptopunk
"ETH got cheaper" 🙈 - good times
Got my first #cryptopunk 5543
Looked like cowboy bebop. ETH got cheaper. Drinks were involved. 🥂 Here we are
#NewProfilePic #cryptopunks
4
561
AGI estimates for 2027.
Ethereum 3.0 for 2029.
So we might see it in 2028?
I am sorry but does this actually say Ethereum 3.0 ships in 2029
Nothing ships on time so realistically it'll be 2031
THAT IS 7 YEARS FROM NOW
And that's just to improve the consensus layer (not DA nor the EVM)
That is actual insanity. Go ship a damn product.
2
494