Must be nice to have unquota’ed tokens to use with frontier AI (is this the case for Anthropic employees?). One thing I think is fascinating as we enter the Intellicene is the disproportionate access to AI. The ability to petition them to do what you want is currently based on monthly subscriptions, but will it change in the future? Who knows?
I don't understand how these setups scale longterm, and even more so for the average user. The latter is relevant because, as he points out, his setup isn't that far out of reach of the average person - it's still fairly close to out of the box claude code, and opus.
But between the model qualities varying, the pricing, the timing, the tools constantly changing, I think it's really difficult to build the institutional knowledge and setup that can be used beyond a few weeks.
In the era of AI, I don't tink it's good enough to "have" a working product. It's also important to have all the other things that make a project way more productive, like stellar documentation, better abstractions, clearer architecture. In terms of AI, there's gotta be something better than just a markdown file with random notes. Like what happens when an agent does something because it's picking something up from some random slack convo, or some minor note in a 10k claude.md file. It just seems like the wild west where basic ideas like additional surface area being a liability is ignored because we're too early in the cycle.
tl;dr If it's just pushing around typical mid-level code, then... I just think that's falling behind.
I doubt he’d use Claude code as it is. I’m sure he’d upgrade to think harder and do more iterations and go deeper. Codex for example already does that but could go deeper a bit longer to figure out more.
I'm a bit jealous. I would like to experiment with having a similar setup, but 10x Opus 4.5 running practically non stop must amount to a very high inference bill. Is it really worth the output?
From experimentation, I need to coach the models quite closely in order to get enough value. Letting it loose only works when I've given very specific instructions. But I'm using Codex and Clai, perhaps Claude code is better.
Absolutely shocking... Boris uses a light themed terminal?! Kidding aside, these were great tips. I am quite intrigued by the handing off of local Claude sessions to the web version. I wonder if this feature exists for the other Coding CLI agents.
This was extremely useful to read for many reasons, but my favorite thing I learned is that you can “teleport” a task FROM the local Claude Code to Claude Code on the web by prepending your request with “&”. That makes it a “background” task, which I initially erroneously thought was a local background task. Turns out it sends the task and conversation history up to the web version. This allows you to do work in other branches on Claude Code web, (and then teleport those sessions back down to local later if you wish)
I implemented some of his setup and have been loving it so far.
My current workflow is typically 3-5 Claude Codes in parallel
- Shallow clone, plan mode back and forth until I get the spec down, hand off to subagent to write a plan.md
- Ralph Wiggum Claude using plan.md and skills until PR passes tests, CI/CD, auto-responds to greptile reviews, prepares the PR for me to review
- Back and forth with Claude for any incremental changes or fixes
- Playwright MCP for Claude to view the browser for frontend
I still always comb through the PRs and double check everything including local testing, which is definitely the bottleneck in my dev cycles, but I'll typically have 2-4 PRs lined up ready for me at any moment.
1) everyone on the team uses Claude code differently.
2) Claude Code has been around for almost a year and is being built by an entire team, yet doesn't seem to have benefited from this approach. The program is becoming buggier and less reliable over time, and development speed seems indistinguishable from anything else.
3) Everything this person says should be taken with a massive grain of salt considering their various conflicts of interest.
> The program is becoming buggier and less reliable over time
You know, this bugs me out. Claude code, macOS, Windows... they're all becoming buggy, filled with papercuts everywhere... and this coincided with layoffs + mass adoption of LLM coding tools and services.
The funniest part of that whole thing was when someone said "I trusted you, but you use light mode on your terminal" and then he replied that people stop by his desk daily just to make fun of him for it.
I actually use dozens of claude codes "in parallel" myself (most are sitting idle for a lot of the time though). I set up a web interface and then made it usable by others at clodhost.com if anybody wants to try it (free)!
Having the 5 instances going at once sounds like Google Antigravity.
I haven't used Claude Code too much. One snag I found is the tendency when running into snags to fix them incorrectly by rolling back to older versions of things. I think it would benefit from an MCP server for say Maven Central. Likewise it should prefer to generate code using things like project scaffolding tooling whenever possible.
Frankly Claude code is painfully slow. To the point I get frustrated.
On large codebases I often find it taking 20+ minutes to do basic things like writing tests.
Way too often people are like it takes 2 minutes for it to do a full pr. Yeah how big is the code base actually.
I also have a coworker who is about 10x more then everyone else. Burning through credits yet he is one of the lowest performers.{closing in on around 1k worth of credits a day now).
This feels like the desperate, look at me! post, which is the exact opposite of Andrej Karpathy's recent tweet[0] about feeling left behind as a programmer, as covered on Hacker News[1].
I guess would want to see how sustainable this 5 parallel AI effort is, and are there demonstrably positive outcomes. There are plenty of "I one-shotted this" examples of something that already (mostly) existed, which are very impressive in their own right, but I haven't seen a lot of truly novel creations.
This is interesting to hear, but I don't understand how this workflow actually works.
I don't need 10 parallel agents making 50-100 PRs a week, I need 1 agent that successfully solves the most important problem.
I don't understand how you can generate requirements quicky enough to have 10 parallel agents chewing away at meaningful work. I don't understand how you can have any meaningful supervising role over 10 things at once given the limits of human working memory.
It's like someone is claiming they unlocked ultimate productivity by washing dishes, in parallel with doing laundry, and cleaning their house.
Likely I am missing something. This is just my gut reaction as someone who has definitely not mastered using agents. Would love to hear from anyone that has a similar workflow where there is high parallelism.
Agree. People are stuck applying the "agent" = "employee" analogy and think they are more productive by having a team/company of agents. Unless you've perfectly spec'ed and detailed multiple projects up front, the speed of a single agent shouldn't be the bottleneck.
The only way to achieve that level of parallelism is by not knowing what you are doing or the peoblem space you are working in to begin with and just throwing multiple ill defined queries at agents until something "works". It's sort of a modern infinite monkey theorem if you will.
It would be very interesting to see the outputs of his operations. How productive is one of his agents? How long does it take to complete a task, and how often does it require steering?
I'm a bit of a skeptic. Claude Code is good, but I've had varied results during my usage. Even just 5 minutes ago, I asked CC to view the most recent commit diff using git show. Even when I provided the command, it was doing dumb shit like git show --stat and then running wc for some reason...
I've been working on something called postkit[1], which has required me to build incrementally on a codebase that started from nothing and has now grown quite a lot. As it's grown, Claude Code's performance has definitely dipped.
I spent a whole day running 3x local CC sessions and about 7 Claude code web sessions over the day. This was the most heavy usage day ever for me, about 30 pull requests created and merged over 3 projects.
I got a lot done, but my brain was fried after that. Like wired but totally exhausted.
Has anyone else experienced this and did you find strategies to help (or find that it gets easier)?
61 comments
[ 130 ms ] story [ 2052 ms ] threadIt's interesting that Boris doesn't mention "Agent Skills" at all. I'm still a bit confused at the difference between slash commands and Agent Skills.
https://code.claude.com/docs/en/skills
But between the model qualities varying, the pricing, the timing, the tools constantly changing, I think it's really difficult to build the institutional knowledge and setup that can be used beyond a few weeks.
In the era of AI, I don't tink it's good enough to "have" a working product. It's also important to have all the other things that make a project way more productive, like stellar documentation, better abstractions, clearer architecture. In terms of AI, there's gotta be something better than just a markdown file with random notes. Like what happens when an agent does something because it's picking something up from some random slack convo, or some minor note in a 10k claude.md file. It just seems like the wild west where basic ideas like additional surface area being a liability is ignored because we're too early in the cycle.
tl;dr If it's just pushing around typical mid-level code, then... I just think that's falling behind.
It offers a lot more context and I found it more helpful than the original twitter thread
From experimentation, I need to coach the models quite closely in order to get enough value. Letting it loose only works when I've given very specific instructions. But I'm using Codex and Clai, perhaps Claude code is better.
My current workflow is typically 3-5 Claude Codes in parallel
- Shallow clone, plan mode back and forth until I get the spec down, hand off to subagent to write a plan.md
- Ralph Wiggum Claude using plan.md and skills until PR passes tests, CI/CD, auto-responds to greptile reviews, prepares the PR for me to review
- Back and forth with Claude for any incremental changes or fixes
- Playwright MCP for Claude to view the browser for frontend
I still always comb through the PRs and double check everything including local testing, which is definitely the bottleneck in my dev cycles, but I'll typically have 2-4 PRs lined up ready for me at any moment.
1) everyone on the team uses Claude code differently.
2) Claude Code has been around for almost a year and is being built by an entire team, yet doesn't seem to have benefited from this approach. The program is becoming buggier and less reliable over time, and development speed seems indistinguishable from anything else.
3) Everything this person says should be taken with a massive grain of salt considering their various conflicts of interest.
You know, this bugs me out. Claude code, macOS, Windows... they're all becoming buggy, filled with papercuts everywhere... and this coincided with layoffs + mass adoption of LLM coding tools and services.
I haven't used Claude Code too much. One snag I found is the tendency when running into snags to fix them incorrectly by rolling back to older versions of things. I think it would benefit from an MCP server for say Maven Central. Likewise it should prefer to generate code using things like project scaffolding tooling whenever possible.
On large codebases I often find it taking 20+ minutes to do basic things like writing tests.
Way too often people are like it takes 2 minutes for it to do a full pr. Yeah how big is the code base actually.
I also have a coworker who is about 10x more then everyone else. Burning through credits yet he is one of the lowest performers.{closing in on around 1k worth of credits a day now).
but also
> Claude Code works great out of the box, so I personally don't customize it much.
Am I the only one to notice the irony of this juxtaposition?
I guess would want to see how sustainable this 5 parallel AI effort is, and are there demonstrably positive outcomes. There are plenty of "I one-shotted this" examples of something that already (mostly) existed, which are very impressive in their own right, but I haven't seen a lot of truly novel creations.
[0] https://x.com/karpathy/status/2004607146781278521
[1] https://news.ycombinator.com/item?id=46395714
I don't need 10 parallel agents making 50-100 PRs a week, I need 1 agent that successfully solves the most important problem.
I don't understand how you can generate requirements quicky enough to have 10 parallel agents chewing away at meaningful work. I don't understand how you can have any meaningful supervising role over 10 things at once given the limits of human working memory.
It's like someone is claiming they unlocked ultimate productivity by washing dishes, in parallel with doing laundry, and cleaning their house.
Likely I am missing something. This is just my gut reaction as someone who has definitely not mastered using agents. Would love to hear from anyone that has a similar workflow where there is high parallelism.
I'm a bit of a skeptic. Claude Code is good, but I've had varied results during my usage. Even just 5 minutes ago, I asked CC to view the most recent commit diff using git show. Even when I provided the command, it was doing dumb shit like git show --stat and then running wc for some reason...
I've been working on something called postkit[1], which has required me to build incrementally on a codebase that started from nothing and has now grown quite a lot. As it's grown, Claude Code's performance has definitely dipped.
[1] https://github.com/varunchopra/postkit
I got a lot done, but my brain was fried after that. Like wired but totally exhausted.
Has anyone else experienced this and did you find strategies to help (or find that it gets easier)?