Twenty agent tabs is not a workflow. One owner inbox might be
Start a task, the agent is slow, start another. The first one replies and wants a clarification, a follow-up and a side task, each in its own tab. Twenty threads later your head is the only place where the state lives, and it's full. Sergey Kuhne calls it AI brain fry, and the Nov 8 post on how they fight it drew 94 points and 80 comments on Hacker News on Sunday.
The fix is a queue, and the queue lives in git. Every ticket is a markdown file in the repo next to the code it's about, and the state of the board is a committed snapshot. Nothing important lives in a chat window or in anyone's memory. If a session dies, the queue is still there.
An agent works that board in cycles. It researches new tickets, builds the ones a human has tagged safe to build, runs tests and a browser check, and verifies what's already deployed. Sensitive code goes on a branch and waits for approval. It never deploys to production.
What it can't do is decide. So a script reads the board and builds one prioritized "owner inbox" of exactly the items waiting on human judgment: review this change, answer this question, rule on this trade-off. The reading surface is a Claude artifact rebuilt from git every cycle, one page per ticket. You open a link, read what was built and why, give a verdict, and the verdict goes back into git. The trick, per the post, is the closing: once you rule on an item, its context leaves your head with it.
There are no metrics here, only one person's setup, and the post admits the hard parts are still loading context fast and keeping blocked tickets from looping. But it's the clearest answer yet to a question users keep asking: where does the human sit once agents run in parallel? Not in twenty tabs. In one inbox with only the decisions that are actually yours.
Source: https://sergeykuhne.com/blog/fighting-ai-brain-fry
← Back to all articles
The fix is a queue, and the queue lives in git. Every ticket is a markdown file in the repo next to the code it's about, and the state of the board is a committed snapshot. Nothing important lives in a chat window or in anyone's memory. If a session dies, the queue is still there.
An agent works that board in cycles. It researches new tickets, builds the ones a human has tagged safe to build, runs tests and a browser check, and verifies what's already deployed. Sensitive code goes on a branch and waits for approval. It never deploys to production.
What it can't do is decide. So a script reads the board and builds one prioritized "owner inbox" of exactly the items waiting on human judgment: review this change, answer this question, rule on this trade-off. The reading surface is a Claude artifact rebuilt from git every cycle, one page per ticket. You open a link, read what was built and why, give a verdict, and the verdict goes back into git. The trick, per the post, is the closing: once you rule on an item, its context leaves your head with it.
There are no metrics here, only one person's setup, and the post admits the hard parts are still loading context fast and keeping blocked tickets from looping. But it's the clearest answer yet to a question users keep asking: where does the human sit once agents run in parallel? Not in twenty tabs. In one inbox with only the decisions that are actually yours.
Source: https://sergeykuhne.com/blog/fighting-ai-brain-fry
Comments