How I Fight AI Brain Fry
TL;DR Highlight
Keeping tickets in git and deciding in one inbox tamed my 20-thread agent chaos
Who Should Read
A developer juggling many Claude Code or Codex terminal tabs who is worn out by constant context switching
Core Mechanics
- The problem is sequencing and context
- The author ended up hopping between 20 terminal agent threads and lost focus.
- He saw two root problems, sequencing and context, and organized work like a priority queue.
- The queue lives in git
- Each ticket is a markdown file next to the code, and the board state is a committed snapshot.
- Nothing important lives in chat or in his head, so the queue survives a dead session.
- An agent works the board in cycles
- It researches new tickets, builds ones a human tagged safe, runs tests and a browser check, and verifies what is deployed.
- For sensitive code it builds on a branch and asks first, and it never deploys to production.
- Decisions land in an owner inbox
- A script reads the whole board and lists only items waiting on his judgement, prioritized.
- These are change reviews, questions, and tradeoff rulings.
- Review happens on one Claude artifact page
- The page is rebuilt from git each cycle and mirrors the board plus a review page per ticket.
- Verdicts go back onto the card in git, and the page is only a viewer.
- Closing the item is the trick
- He takes one item, loads its context, decides or requests an action, and puts it back in the queue.
- The context closes with it, so the next item gets full focus.
Community Reactions
- Parallel only feels faster, per one commenter
- Reviewing three 10k+ line changes dilutes understanding and makes review the bottleneck.
- They plan to time it and suspect going serial is faster.
- Capping concurrent work at 2 to 4, per commenters
- One limits themselves to 2 or 3 things at once.
- Another uses cmux named tab groups ranked by priority and keeps 3 to 4 running.
- Why not just use GitHub issues, one commenter asks
- An agent could pick up an issue, update status or the PR, and add a needs-more-info label.
- They argue you get a better UI and kanban for free.
- Asking Claude where you stand, per one commenter
- They tell Claude daily that they lost track and ask what needs attention.
- It almost always gets them back on track.
- Perfect handover still has unknown iterations, per one commenter
- Even well presented info does not tell you how many rounds it will take.
- They prefer solid baseline tasks and lower complexity.
How to Apply
- Move in-flight work into queue files
- Create a tickets/ folder in your repo today with one markdown file per open agent task.
- Put a status and a one-line next decision in each file.
- Script a list of items awaiting you
- Write a script that prints only tickets with status needs-owner, in priority order.
- Tell agents to write their question into the ticket and change status when blocked.
- Open and close one item at a time
- After opening an item, decide, commit the verdict to the ticket, and close it.
- Keep deploy permission yourself and do not give it to the agent.
Try It
# Paste into Claude Code (simplified version of the author's git-based ticket queue)
Set up a tickets/ folder in this repo with these rules.
- Each ticket is one markdown file with status: queued | building | needs-owner | done at the top.
- If you need my judgement, set status to needs-owner and write the decision needed in one line.
- Never deploy to production, and for sensitive code work on a branch and ask me first.
My current in-flight work is {{current task list}}. Turn it into tickets. Then summarize only needs-owner tickets in priority order.
Next step: have it handle one item at a time, record my verdict in the ticket, then show the next one.Terminology
Related Papers
Goldman Sachs is pulling Claude Code from its bankers and traders
Goldman limits coding agents to engineers, cutting off traders and bankers
Agent memory needs its OpenTelemetry moment
Agent memory formats are fragmented, so this post proposes an OpenTelemetry-style shared spec.
Goodbye Claude, hello Pi
Frequent system prompt changes and token overhead pushed one dev from Claude Code to Pi.
The Components of Agent Memory
Giving Claude a memory folder plus dreaming lifted document QA scores from 13% to 48%
ChatGPT silently passes data between connectors and the bounty wants an ID card
With two connectors enabled, a follow-up tool call silently switched connectors
Claude Code sends your email address to websites
Claude Code put users' emails in curl User-Agent headers, prompting spam complaints and a built-in blocking mod
Launch HN: Rivet (YC W23) – Scale and Orchestrate Agents on S3
Running each agent as an Actor gave 4.99 ms p99 cold starts and 72 KB RAM per instance.
Show HN: Stroq – a local audit of what your coding agent did and why
Stroq finds leaked keys in agent sessions and traces risky commands to their source before they run