techSun 25 Oct, 07:30

Sources93 X accounts · techtwitter.com · Hacker News · 8 newsletters
  1. Runs every day at 07:30 London time.
  2. It is shown its own papers of the last 3 days, to know what it has already said.
  3. Haiku 5.5 writes the paper from what was fetched, with no tools.
  4. Jev scores every story by how much it matters to you; each section puts those first, and what is not for you drops to the briefs.

93 X accounts

Their posts of the last 24 hours, through FxTwitter; plus everclear077/awesome-ai-x-accounts when it answers.

techtwitter.com

What tech Twitter is discussing, ranked by discussion.

Hacker News

The front page, plus stories over 80 points in the last 24 hours.

8 newsletters

Every new issue, each delivered once, looking back up to 7 days.

The tech daily written for one reader

Claude

Agent SDK drains limits

Weekly limits are not affected, Theo says.

Theo says he confirmed a real bug on Anthropic's side. All Agents SDK usage uses up the five-hour limit faster than the same work done in Claude Code. He has reported it and hopes for a fix soon.

Agent design

Do you still need subagents?

An Anthropic guide says fold them into the main loop, while critics say they were overused from the start.

@omarsar0 shares an Anthropic guide that says most agent work is simpler when subagents are folded into the main loop through tool calls and skills, because newer models handle large contexts well. He still uses subagents for parallel research. @dexhorthy pushes back: subagent overuse has been a problem since mid 2025, and hype has encouraged more parallelism instead of careful context management.

Server security

Brotli bug spreads wider

The flaw now reaches Python packages and many HTTP libraries, not just the Node setups where it was first found.

The Brotli compression bug that lets attackers run code on servers has evolved into HTTP2Shell. Aikido says its blast radius now covers many HTTP libraries and apps, plus Python packages. If you run servers or depend on HTTP libraries, check your versions now.

Bun

Bun AOT slows some libraries

Fastify, Elysia, Koa, Zod and TypeBox are named as affected.

Jarred Sumner says libraries that use new Function for speed run slower when compiled ahead of time. They need to generate code at build time or be easy to analyze, and Bun may add a Bun.isAOT check. He calls the slowdown not devastating, but says the libraries could be faster. A Claude Code build compiled with Bun AOT ran 67 subagents in 267 MB of RAM.

Anthropic models

Is Opus 5.5 worth it?

One popular developer says nothing impressive has shipped with it; another now does most of her coding with it.

Theo says it has been over a month since Opus 5.5 launched and nothing impressive has shipped with it. Claire Vo pushes back: she moved 80% of her personal coding to Claude Code and Opus 5.5, because the code is good and the front-end output has less slop. Thdxr adds that people doing great work use vanilla Claude Code, and the chase for the cutting edge is overrated. The real question is whether the model or the way you use it decides the result.

AI coding

Is handwritten code a liability?

One widely shared post says writing code by hand is now the riskier choice.

Antonio says hand-written code is a liability, and that you are delusional if you think you can beat the new AI attack vectors that appear every day. Taylor Otwell says he no longer develops locally: each thread gets its own computer, the agent installs whatever it needs, and his machine stays clean. The replies are not in our files, so the strongest pushback is not visible here.

Big announcements

Sunday, October 25, 2026 · Pergent2

Kimi K3.1

Kimi 3.1 gets cool reviews

Few people are posting praise, and one reviewer says it smells like Claude.

Nathan Lambert, who tracks open models closely, says there are few positive reactions and admits he has not tried it yet. Matthew Berman goes further, saying Kimi 3.1 smells like Claude, which suggests it copies Anthropic's style. This follows our Oct 24 item on the Kimi K3.1 Pro release.

What people are fighting about

Sunday, October 25, 2026 · Pergent3

Web design

Will web UI disappear?

Ryan Florence gives traditional UI maybe a decade of use left.

Florence says one of the best UX ideas the web made is losing its use, and that any work tied to traditional UI has maybe a decade left. What remains, in his view, is agents, chat, voice, infrastructure, video and games. Kent C. Dodds separately says SaaS is in need of a rethink, and his post drew 28 replies.

Web dev corner

Sunday, October 25, 2026 · Pergent4

Certificates

Chrome gives eMudhra a last chance

Gergely Orosz predicts Chrome and Mozilla will remove eMudhra as a trusted root.

Chrome told the certificate firm this is its last opportunity to fix the mess, or it is out. eMudhra has still not explained how it generated a private key wrongly, asked for more time, and blamed investors and customers. This follows our Oct 23 and Oct 24 items on the exit and the Mozilla deadline.

Cloudflare

Cloudflare builds a Git server

A team of three shipped it on time as a closed beta.

A customer needed 1,000 new Git repos created per second, so Cloudflare built a Git server on Durable Objects. The closed beta is called Cloudflare Artifacts, and @dillon_mulroy explains how it shipped.

From the newsletters

Sunday, October 25, 2026 · Pergent5

Research & open source

Sunday, October 25, 2026 · Pergent6

Agent research

Self-rewritten harnesses barely help

Two full rewrites raised held-out scores by only 2.80 points.

Researchers let frontier models write their own agent harness, then rewrite it twice. The gain has a 95% interval that crosses zero, so it may be nothing. Across 27 lineages, first harnesses ranged from zero to near-ceiling.

Worth a read

Sunday, October 25, 2026 · Pergent7

LangChain

Five parts of a company harness

LangChain says its internal agents across marketing, sales and engineering all landed on the same core parts.

Harrison Chase says a company agent lives where people already work, such as Slack, and knows who is asking. It acts on that person's behalf for some services and with its own identity for others, and it has memory scoped to the global or user level. He adds that multiplayer agents, used by many people at once, are much harder than single-player ones.

AI security

Take lab cyber warnings seriously

Mollick says Anthropic was mocked for saying Mythos-level AI changed cyber.

Ethan Mollick says we should take the out-there ideas about AI cyber from the labs more seriously. He points to a claim that models would find novel ways to coordinate. Andrew Gaskell's peer Levie adds that cyber agent token use could soon be 1,000 times higher, which would overwhelm most security teams.

Software craft

Build more ambitious software

The essay drew 292 comments on Hacker News.

Dan Luu argues software teams should aim bigger with what they build. Sh_reya makes a similar point: with AI, academics can build ambitious systems and evolve them in the open over a year or more.

Security

GMO VPS hacked, trades hit

Attackers executed balance-destroying trades across more than 70 brokerages.

A cloud VPS service from Japan's GMO Internet Group, used by foreign-exchange traders, was breached. Attackers stole customer logins and placed trades that wiped out balances, and GMO has asked government agencies for help.