Загрузка видео...
Не удалось загрузить видео
I pointed the new Claude Managed Agents dynamic workflows at 18k lines of code. > 21 proven bugs. > 5 minutes. > $8.36, on my Claude Max API credits. A lead agent writes a plan, runs it across many agents in phases, then merges the results. Try it in... show more
11,788 просмотров • 1 месяц назад •via X (Twitter)
Комментарии: 23

This is my Saturday

Enjoy it!

faster usage burn is the key to success 🔥

Guess it depends on your definition of success 😂

now the fun part: finding out how many of the 21 the author calls "by design"

Cheapest bug bounty program ever run lol

Worth every Penny!

What happened with the sol-advisor, going to update it so it runs sol 6.1?

that /claude-api is to die your api access?. if So my mythos 5.1 access is via api so this would be useful

I think it basically gives Claude the context to build with the API

what's the false positive rate? an agent emitting 21 findings in 5 min is cheap; 21 findings that each reproduce with a failing test is a result.

21 bugs in 18k lines for like $8 is wild ngl how many held up once you actually read through them? and did the merge step ever choke when two agents touched the same file

8 bucks to find 21 bugs in five minutes. my code review process is just me squinting at the screen until i feel vaguely guilty.

有点好奇这批 bug 里有几个是现有测试没兜住的

awesome.

thats not too bad - how bad were the bugs? also did it run opus for the agents?

21개 버그가 실제 검수에서 얼마나 남는지와 에이전트 단계별 비용을 같이 기록하면, 병렬 실행의 효율을 더 정확히 비교할 수 있겠네요.

love this for work that's parallelizable and can run in isolation. but dynamic workflows don't solve what i'm still figuring out with 50-100 agents: how to keep them building features and real value without stepping on each other's toes, especially when different people are running them

the interesting bit isn't the agent count, it's the orchestration. 21 real bugs in 18k lines from one automated pass is honestly wild — results like that make the "agent swarms are just token waste" take a lot harder to hold

a lead agent that plans then fans out and merges is the workflow that finally feels like it scales past one chat window

It’s like sending a lead engineer to scout the codebase while specialists inspect separate areas, then reconciling the findings into one review.

Use your own agents outside of Claude Code. They can collaborate and share permission-based access to files behind your own Blackboard.

Does each finding come with a runnable repro or regression test? Would be useful to see one of the bugs from discovery through to a verified fix.




![What’s the most you’ve ever spent on Claude Code? 👀 Claude is still one of my favorite AI coding tools for frontend work, but this is getting ridiculous. Across these sessions, the API estimate came out to $744.61 for Claude Code vs $20.55 for Codex Claude accounted for 97.3% of the total estimated cost. I understand paying more for a better model, especially when the frontend quality is there. But at some point you have to ask whether the extra cost and usage hurdles are actually worth it for everyday development. Claude is incredible. Claude pricing is also incredible… just in a completely different way 😭 What’s the most you’ve spent on Claude Code? [Claude Code, Anthropic, Codex, AI coding, AI agents, vibe coding, coding agents]#greenscreen](https://ttanon.com/i/c/7679991999654645022-720.webp)