惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

I
InfoQ
G
Google Developers Blog
Engineering at Meta
Engineering at Meta
月光博客
月光博客
博客园 - 聂微东
博客园 - Franky
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
有赞技术团队
有赞技术团队
A
About on SuperTechFans
Microsoft Azure Blog
Microsoft Azure Blog
Blog — PlanetScale
Blog — PlanetScale
U
Unit 42
T
Tailwind CSS Blog
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
云风的 BLOG
云风的 BLOG
S
SegmentFault 最新的问题
F
Fortinet All Blogs
H
Help Net Security
J
Java Code Geeks
酷 壳 – CoolShell
酷 壳 – CoolShell
博客园 - 叶小钗
L
LangChain Blog
Martin Fowler
Martin Fowler
N
Netflix TechBlog - Medium

Hacker News: Ask HN

The New Window Delete ChatGPT Atlas Spyware Tell HN: Qwen Free Tier Is Discontinued Ask HN: SeedLegals Partnerships in London, worth it? Ask HN: How to highlight talent from untraditional backgrounds? Ask HN: We dont need a programming language now? Durable Object alarm loop: $34k in 8 days, zero users, no platform warning What if Time at the subatomic level has multiple arrows? How to add MidnightBSD Key to UEFI Secure Boot DBX? (Revoked and Forbidden Keys) Ask HN: What's your experience working at xAI as an AI tutor? Any engineers here with experience of clinical data standards? Ask HN: Who is using OpenClaw? Agent Skills for Software Test Automation Ask HN: Who needs contributors? Claude Code is thinking too much Ask HN: What Is the Big-O Order of a Jigsaw Puzzle? Ask HN: Stepping into a new role as a Senior, mentoring dos and dont's? Founder from Zurich heading to SF and Austin for the first time Hacker News No Manual Screenshots: I Built a Scalable Screenshot API Using Cloud Playwright Ask HN: Thought experiment: AGI giving us answers we don't like? Ask HN: I quit my job over weaponized robots to start my own venture 1% Vacancy, 81% Preleased: Where Midmarket Compute Deploys in 2026 Ask HN: Preferred pricing model for sound effects libraries? Copy of the email I sent to my undergraduate professors on Nov 30, 2025 Model API Performance | Hacker News Ask HN: Are open-weight LLMs the new offline encyclopedias? Valgrind 3.27 RC1 is out Claude Code OAuth down for >12 hours Ask HN: What's Better?–Tauri or Electron?
Understanding how to run a fleet of agents
sermakarevic · 2026-05-20 · via Hacker News: Ask HN

I started with materializing replies as markdown artifacts. Anything that stays in the terminal gets lost and forgotten quickly, so persisting each reply to a file was the first step toward keeping the work durable.

Once the number of artifacts grew, a flat folder stopped being usable. I introduced hierarchical organization — grouping by topic, with a top-level index.md pointing to per-topic index.md files, each of which holds a short summary and links to individual artifacts. Nesting depth is unbounded; you add a level whenever a topic earns one.

The next realization was that tasks belong in a queue, not to the terminal. I moved to markdown files with todo and done sections. At that point, almost no terminal interaction was left — just /clear after each task and a prompt like "take the next task."

The obvious next step was to automate that loop: a shell loop calling claude -p "take next task", with context cleared automatically at the end of each session. This works fine for a single worker.

Scaling beyond one agent needs a few extra pieces: - A proper queue. This is what beads provides — dependency-aware, atomic claims via bd ready and bd update --claim, so two workers don't grab the same task. - A channel back to the human. Also handled by beads: bd update <id> --status blocked --notes "QUESTION: <one-line question>" parks the task and surfaces the question without halting the rest of the fleet. - Session memory. Because each claude -p invocation starts with a fresh context, the agent has to write down what it has done, what it plans to do, and any related notes — either into beads itself (status, notes, dependencies on the issue) or to disk — so the next session can pick up cleanly after a restart.

With these components in place, we can spawn many claude workers pulling from tasks from beads db, asking questions, adding tasks to db, storing progress and final artifacts to disk. AMD describe this pattern in their amazing issues description https://github.com/anthropics/claude-code/issues/42796. It took me some time to figure out how this suppose to work.