What if a codebase was actually stored in Postgres and agents directly modified files by reading/writing to the DB?
Code velocity has increased 3-5x. This will undoubtedly continue. PR review has already become a bottleneck for high output teams.
Codebase checked-out on filesystem seems like a terrible primitive when you have 10-100-1000 agents writing code.
Code is now high velocity data and should be modeled at such. Bare minimum, we need write-level atomicity and better coordination across agents, better synchronization primitives for subscribing to codebase state changes and real-time time file-level code lint/fmt/review.
The current ~20 year old paradigm of git checkout/branch/push/pr/review/rebase ended Jan 2026. We need an entirely new foundational system for writing code if we’re really going to keep pace with scale laws.
This is diametrically opposed to my view. I prefer to use a single language to build a distributed system. Languages like @unisonweb and @ActonLang are the future.
This is why I’m unimpressed by Erlang/Elixir: every major language runtime has VERY high-quality M:N work-stealing “thread” schedulers with good APIs (structured concurrency), and the “isolated processes” and “RPC” got pushed up to an orchestration layer (DC/OS, Nomad, k8s…)
@lemire@grok The 46Gbps of Wifi 7 is theoretical and for the air side. I think in practice (if you even have such new devices) you could get something like 15-20 Gbps in best case (like direct line of sight & few meters to AP), so about 1.5-2GB/s, much lower than the 5 you mentioned.
How is it Codex can consistently write 50 lines of Haskell one-shot but fail to write proper git commit messages. One is ELI5 "50 chars as summary\n\nthen more stuff with-word-wrap" and the other is like years of FP-knowledge. So weird and backwards..
@ntdvps are NETCONF locks supposed to be working on Nokia SR Linux? When I lock candidate, do edit-config & commit, I can only run one commit per NETCONF session, second commit comes back with error, regardless of config payload etc. Not using locks works.
@satnam6502 I'm pretty sure my next laptop will be air again, although it'll be awhile so things might change of course. For extra screen you use usb-c anyway, which the air has. I rather carry hdmi dongle for the few cases than constantly carry the extra weight in the laptop
@DrPhiltill@jgarzik Claude code with Claude 4 models (opus for advanced stuff) is really the best tool for coding. I think it's wise to try it before concluding anything about the current state of AI / vibe coding. Also prompt instruction are still relevant and have a big impact on success rate
I use agents frequently, letting them run wild on my code base. But at the end of the day, I commit and I stand for my work. Sometimes I ask agents for explanations, but regardless I read and understand every line I commit. I don't think you have to restrict an agent to a window.
Possibly the best thoughts on AI usage that resonates with my feelings. I too like to use AI in the small while coding but the main loop has to be driven by me with my fingers. I feel too nervous watching Claude agent going through my codebase and making changes without me going through the cognitive exercise of forming the mental model of the program. The only time I let the agent loose is when I am trying to understand a large codebase and asking the AI to prepare a workflow of how the code works. But that’s only a read mode operation on the codebase. Learning through fingers, as @dhh says is a great idiom to learn programming.
Zero dependencies is awesome but obviously not for your day to day code on average run of the mill programs where reimplementing everything would be madness. Only listen to this if you are building something very serious.
I wish I had the zero dependency discipline Joran and the TigetBeetle folks have because more often than not, dependencies do end up being a liability or that thing that prevents you from engineering the right thing.
@lemire I think the IETF motto of "be conservative in what you send, and liberal in what you accept" has done a lot of harm this way. It's well intentioned but as you say, instead of finding and fixing errors we now have to carry with us loads of special case handling on the receive side