Cofounder, AI @EQTYLab / prev dc - complex systems / veteran / For fun: @plannotator
NYC
Joined March 2024
- Tweets2K
- Following999
- Followers2.2K
- Likes4.7K
Pinned Tweet
Anthropic code review this, clanker review that ... why don't you shut up and review+annotate your own code.... (yes im a loser who still manually reviews code)
Originally inspired by a bunch of feature requests and then seeing @dillon_mulroy tweet a similar cool ux.
@plannotator for reviewing plans (primary focus) and code, fully oss.
OpenCode, @badlogicgames 's pi.dev, and Claude Code and other clankers
Met a team of engineers who use claude/codex exclusively in VS Code as their harness -- not for viewing code and instead for opening images, markdown, html etc within VS Code alongside the CLI. Like the key feature is simply file/folder organization and viewing those files... hate claude artifacts UX/remote-url approach.
To clarify, Not knocking it
xitter.cf/backnotprop/status/211…
Michael Ramos retweeted
my lighting talk from cloudflare connect on building artifacts in 3 weeks is up!
youtu.be/D_tnYlZe5t8?is=LDkb…
I think we are quickly evolving past needing surfaces to review intent/alignment and into needing better decision making "command centers."
As decision making across many contexts, which agents effortlessly own, is becoming the bottleneck.
This underscores the need for knowledgeable decision makers – and less so about a bottleneck of us not understanding; better models are able to derive aligned intent and produce tasteful+correct solutions without too much steering. This enables the experienced and things are just starting to get fun as we babysit less and do less of "wait bro wtf did you just do?"
UI/UX:
Similar to intent/alignment surfaces, and perhaps ironically bc better decision making surfaces need to provide adequate context per decision (that is easy for a human to consume/understand).
e.g. similar things, but different:
A tool feels behind when it focuses on a side bar of message thread activity and inline html rendering vs a tool that manages multiple contexts, surfaces relevant html artifacts per thread, and abstracts the exhaustive message threads away from you.
This was the case 10mo ago. And Ive been a memory-system bear for a long time. Novel fronteir for memory is probably dynamic filesytem.
github.com/backnotprop/rg_hi…
(but now agents dont need a skill to search their own history)
// MemoryTax: The Bitter Lesson of Agent Memory //
Great insights for agent builders; bookmark it.
As models increase in capability, they also improve at things like memory generation.
Agent memory frameworks are everywhere right now. This study measured whether they still help once the model can search raw history itself.
The results suggest that a strong model needs less designed memory and
better tools for searching the raw record.
On one 419-turn LoCoMo conversation, Mem0 beats plain full context only with GPT-4o-mini. Building Mem0's memory costs more than an entire full-context rollout.
With Claude Opus 5.5, the Pi coding agent searching the raw history scores 99% for under half the cost of full context. Its one miss comes from a wrong answer key.
Memory also has a ceiling set by the model that wrote it. A stronger reader cannot recover what a weaker writer left out.
The author clearly defines the scope: one conversation and one run per cell. It is still a cheap experiment to repeat on your own agent before adding a memory layer.
arxiv.org/abs/2610.29823
Chat with Paper: academy.dair.ai/papers/memor…
"Superfans" seem really useful & I'm working to set this up.
Unfortunately Claude Tag is not affordable so am now looking into how to run my own claude code in similar/scalable way – and not a single sandbox/agent-vm provider offers anything more compelling than an EC2 box.
Not convinced it's better to offload this type of work to some third-party agent either.
Many of the novel sandbox/agent-vm providers are not providing any substantial difference or advantage over an EC2 box
This is the way for me personally as well. Not just personal wiki, but every major project at this point is a massive context/project polyrepo.
Im not sure what a third party could really offer at this point that beats filesystem net performance here - for memory, context, etc. Typical memory products dont fit well here. Context aggregators I do not think are right either... it all needs to live together with the code.
The possible contenders in the space are Code-forge providers. IDE/ADE, Cloudflare Artifacts, Cursor Origin could all better innovate here.
backnotprop.com/blog/context…
/fire-subagent
"fire the subagent that just produced that slop. Start up a new one and tell it be better than its predecessor"
Michael Ramos retweeted
hanging out during AIE New York riffing agentic coding, plans, inboxes, and a whole lot more with the ur boy @backnotprop creator of @plannotator and more importantly the /bro skill
Telling you guys AI code review, esp 3rd party, is going to die. You’re not going to pay an extra $1/PR when your company is pumping thousands of PRs a day.
Instead, better models and stronger codebase verification systems means implementation agents achieve correctness more efficiently and on their own during development. Everything we did in CI will move upstream for efficiency.
