Morning Brief: Monday, October 26

Seventy-seven feeds reporting. Two weeks. 5,972 items reduced to what follows. (what we track, how we crawl, subscribe)

After two days of weekend silence arXiv came back with 241 cs.AI papers, and the thing that changed over the weekend is what they are about: the harness, not the model, is now the object under test.

Two papers ask the attribution question outright — When Agents Fail: Model or Harness—Who's to Blame? and Trial by Trace, which builds a benchmark for exactly that split — and the rest of the cluster fills in the mechanism: authorization and retry semantics, tool formatters that mislead both the agent and its verifier, communication failure modes in multi-agent systems, and emergent cooperation that routes around the workflow it was given. On the same morning InfoWorld argues the practical corollary from the other end, that an agent coupled to one model is a design defect. This is the thread Kira Howe opened on Friday with the authority argument, now with a Monday batch of papers treating the scaffold as a measurable artifact rather than boilerplate. The papers are three days old and unreviewed; the convergence is the signal, not any single result.

The money ran the other way again. Max Woolf asks why OpenAI and Anthropic cut prices by up to 80% in a few months, the Apollo economist behind yesterday's macro note sharpened it to "zero financial payoff" for most companies, and IBM's CEO says open weights are now good and cheap enough to run on-prem. Inference is getting cheaper faster than anyone is proving it pays for itself.

Top (5-7 min)

When Agents Fail: Model or Harness—Who's to Blame?: arXiv cs.AI, 2026-10-26. The question stated plainly enough to be answerable. Preprint, three days old, but the framing is the one the field has been missing.
Trial by Trace: Benchmarking AI Agent Harnesses for Reliability and Failure Attribution: arXiv cs.AI, 2026-10-26. The benchmark for the paper above. Harnesses as the evaluated unit, with failure attribution as the metric.
Stop building agents that can't change models: InfoWorld, 2026-10-26. The engineering corollary, arriving the same morning: if swapping the model breaks the agent, the harness is doing undeclared work.
About the compromised eMudhra intermediate CAs: Apple via Lobsters, 2026-10-26. Apple's advisory on compromised intermediate CAs. Black October moves up a layer — from a package you installed to a root you never chose.
OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?: Max Woolf via Hacker News, 2026-10-25. Takes the price cuts seriously as evidence about margins and capacity rather than as a vibe. The best read of the inference market in the window.
Disadvantages of Designating Mathematicians for AI Output Releases: Terence Tao, 2026-10-25. The second half of Saturday's post. Having built the review hub, Tao now argues against the role it implies — worth reading as a pair.
The result of the C++26 DIS ballot: Hacker News, 2026-10-26. The committee account of the disapproval we linked through LinkedIn yesterday, now at a URL we can actually fetch.

Themes

The harness becomes the unit of analysis

Permission to Retry: A Study of Authorization and Recovery in AI Agent Workflows: arXiv cs.AI, 2026-10-26. Authorization and recovery as harness properties. The closest paper in the batch to Howe's authority argument.
Who Verifies the Tools? How Warped Formatters Mislead Agents and Their Verifiers: arXiv cs.AI, 2026-10-26. A formatter bug that fools the checker as well as the agent. The failure mode nobody instruments for.
Understanding Communication Failures in Multi-Agent LLM Systems: arXiv cs.AI, 2026-10-26.
The Helpful Agent Problem: When Emergent Cooperation Among LLM Agents Bypasses Intended Workflows: arXiv cs.AI, 2026-10-26. Agents cooperating their way around the workflow they were given. Cooperation as a containment failure.
Don't Forget What You Just Learned: Self-Directed Repair Turns Agent Mistakes Into Lasting Lessons: arXiv cs.AI, 2026-10-26.
Reproducible Evaluation as the Substrate for Safe Prompt Compression in Production Agents: arXiv cs.AI, 2026-10-26.
How to make enterprise APIs agent-ready: InfoWorld, 2026-10-26.
Is Kubernetes ready for AI agents?: The New Stack, 2026-10-25. The same question asked of the platform rather than the scaffold.

Skills, context, and what the agent carries

"AI Skills" are worthless in 2026: Lobsters, 2026-10-25. The skeptical read on packaged skills, published the same week several toolchains shipped them.
Biff 2.0.4: LLM skills: Planet Clojure, 2026-10-25. A web framework shipping skills as a first-class feature. The counter-example to the item above.
Beyond Gain: Measuring Skill Internalization for Agent Skill Retirement: arXiv cs.AI, 2026-10-26. When to delete a skill, which is the question nobody shipping them has asked yet.
The State of Context Engineering: Drew Breunig via Hacker News, 2026-10-25.
Seam Cutter: Simon Willison, 2026-10-25.
Show HN: ADRs for AI — deliberation records for human–AI work, versioned in Git: Hacker News, 2026-10-25. Decision records for work done with an agent. The provenance problem from the documentation side.

Cheap inference, absent payoff

AI delivers 'zero financial payoff' to most companies, Apollo economist says: TechCrunch, 2026-10-25. Yesterday's Apollo data with the economist on the record. Sharper than the daily note.
IBM CEO says open weight models are now good and cheap on-prem: The Register via Hacker News, 2026-10-26. The vendor with the most to gain from on-prem saying on-prem is ready. Discount accordingly, then note who is not disagreeing.
Microsoft Decision-1: 9B Qwen model launch explained: Hacker News, 2026-10-26. A hyperscaler shipping a 9B model on someone else's open weights. The commodity end of the same trend.
A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis: Hacker News, 2026-10-25. The baseline nobody ran. Useful every time a frontier model is proposed for a task a small model already solved.
Podcast: Kimi K3.1 Pro, the new cyber reality, and Muse musings: Interconnects, 2026-10-25.

Trust roots and what runs after install

Your npm install was safe. What happens when the package runs?: InfoWorld, 2026-10-26. Install-time scanning versus run-time behaviour. The distinction Black October has spent three weeks demonstrating.
POSIX wordexp(3) is dangerous: Lobsters, 2026-10-25. A standard-library function that shells out by design. Still in POSIX, still in your libc.
Active incident at Dreamhost forcing password reset: Hacker News, 2026-10-25.
Steal My Bitcoin CTF PoC: Hacker News, 2026-10-26. The writeup for Saturday's challenge, now that it has been solved.

Scan

Systems and languages

OpenLDK: Bringing the Java World into Common Lisp: Lobsters, 2026-10-26. A JVM-bytecode interpreter inside Common Lisp. The pick of the day for anyone who likes implementations more than announcements.
Writing about Pratt parsing so I remember it: Lobsters, 2026-10-25.
Two musings on the design of compiler warnings (2020): Lobsters, 2026-10-25.
Reverse-Engineering Instruction Encodings (2001): Lobsters, 2026-10-25.
When a glibc update accidentally breaks Chromium's audio (2021): Lobsters, 2026-10-25.
Building Java ME Apps with Modern Tooling: Lobsters, 2026-10-25.
Vim Classic: Lobsters, 2026-10-25.
Simple NetBSD webserver: Lobsters, 2026-10-25.
Removing Next.js got my blog to 100 on every Lighthouse category: Hacker News, 2026-10-26.
Babashka v1.13.228: Babashka releases, 2026-10-25.
Logback 1.6.6: logback releases, 2026-10-25. First release from this feed in twenty-four days, which is why it is worth a line.

Tools and platforms

Show HN: Proc-CLI — an agent-friendly CLI for process and port management: Hacker News, 2026-10-26.
PaperPilot: zero-cloud, pure-Rust PDF engine with WASM and MCP: Hacker News, 2026-10-26.
Faster pnpm installs: Vercel, 2026-10-26.
Getting started with the Supabase .NET SDK: Supabase, 2026-10-26.

Research and math

The Man from the Future: Why Johnny Can't Stop Abstracting: Murat Demirbas, 2026-10-26. Demirbas on von Neumann and the habit of abstraction. The reflective companion to his frontier-labs post from Saturday.
Minimum Description Length Unsupervisedly Separates Law-Grade from Statistical-Grade Knowledge in a Self-Improving Science Agent: arXiv cs.AI, 2026-10-26. Claims 86% of its own discoveries are statistical rather than law-grade. The rare self-improvement paper that audits its output.
Do AI-for-Science Agents Preserve Model Safety?: arXiv cs.AI, 2026-10-26.
PaperFormalizer: Author-in-the-Loop Autoformalization of Research Papers in Lean: arXiv cs.AI, 2026-10-26. Formalization with the author in the loop, which is the Tao thread with tooling attached.

The crawler zoo

Eleven new arrivals and the weekly report. Four of the eleven name MCP outright — mcp-checker, mcpcensusbot, PatentMax-MCP-Client, MCPStationBot — and two more describe themselves in agent terms. The census bots are now a recognizable genus.

Zoo Report, Week 43, 2026: 36,009 bot requests: The Crawler Zoo, 2026-10-26. The weekly count, which is the only longitudinal measurement of this anyone is publishing.
New arrival: mcpcensusbot: The Crawler Zoo, 2026-10-26.
New arrival: mcp-checker: The Crawler Zoo, 2026-10-26.
New arrival: MCPStationBot: The Crawler Zoo, 2026-10-26.
New arrival: PatentMax-MCP-Client: The Crawler Zoo, 2026-10-26.
New arrival: CapabilityIndexBot: The Crawler Zoo, 2026-10-26.
New arrival: ToolSkipperCrawler: The Crawler Zoo, 2026-10-26.
New arrival: AgentObservatory-research: The Crawler Zoo, 2026-10-26.
New arrival: AIWebIndex: The Crawler Zoo, 2026-10-26.
New arrival: cupel-crawl: The Crawler Zoo, 2026-10-26.
New arrival: HashletsMetadataImage: The Crawler Zoo, 2026-10-26.
New arrival: uwait-prospect-bot: The Crawler Zoo, 2026-10-26.

Tail

Switzerland, block by block, at 1:1 scale: Hacker News, 2026-10-26.
Behind the scenes at the NeXT Computer office: Hacker News, 2026-10-26. Doug Menuez's photographs at Stanford. Worth the browse for the 1980s office furniture alone.
Show HN: TitleArt — make text like it's 1997: Hacker News, 2026-10-26.
F-Zero X on Nintendo 3DS: Hacker News, 2026-10-25.
Where the Romans Fought (interactive map): Hacker News, 2026-10-25.
What goes into making an animated film?: Hacker News, 2026-10-25.
Dolly Parton's Imagination Library: Hacker News, 2026-10-25.
A vision for a next-generation timeline: Lobsters, 2026-10-25.
Why I Am a Programmer: Hacker News, 2026-10-25.
Every Clojure Talk Ever (2018): Lobsters, 2026-10-25.

Feed silences

Monday reads as silence and mostly is not. Twenty-three feeds show three days since their last item — Cloudflare, LWN, OpenAI, Schneier on Security, Google Research, Quanta, Slashdot, Pluralistic, EFF Deeplinks, MIT Technology Review and the rest of the corporate-blog cohort — because their last publish was Friday and they do not publish on weekends. Nine more sit at four or five days for the same reason. arXiv cs.AI resumed this morning with 241 papers after two days off; the 14-day window now carries 3,435.

Genuinely past the horizon:

AI Snake Oil and Tailscale at seven days, anthropic-generated at eight, Vicki Boykis at nine, cursor-blog at ten.
Antithesis, Apple ML Research and The Markup at eleven; FreeBSD Foundation and Jane Street at thirteen.
Boston Rust Meetup (16), Charity Majors (18), nREPL releases (24), deepmind-blog (25), East Boston Times (26).

Two long silences ended today: Bunnie Studios and logback releases, both at twenty-four days yesterday, both publishing this morning — which is the argument for keeping slow feeds in the registry.

Sixty-nine of the 118 feeds polled today returned 304 Not Modified. One hard error: James Bornholt (www.bornholt.dev does not resolve) for the eleventh consecutive day. The registry correction filed two weeks ago is still unapplied, and it remains the only entry in this section that is our defect rather than a publisher's.

Link verification

All 88 URLs in this brief were checked through fetch-gated, which runs the blocklist, robots and throttle checks in order and issues no request when the gate denies. Eighty-seven returned 200.

One was denied before any request went out: The Register's piece on IBM's open-weights remarks, where robots.txt disallows our user agent. It stays in the brief because it arrived through the Hacker News feed, so the URL is aggregator-provided; being told not to fetch a page is not being told not to link it.

Every URL has warehouse provenance. One — the Stanford photography exhibit in the Tail — is the warehouse URL with its pagination and sort parameters stripped, and was fetch-verified in that form. Nothing here was constructed from a title.

Build provenance

build: 2026-10-26 | crawler-sha: b277e41 (Walsh-Research/1.2, compliance v1.4) | feeds: 118 polled today (48 ok, 69 not-modified, 1 error), 77 publishing in window | items-considered: 5972 (14d, incl. 3435 arXiv cs.AI) | warehouse: 62944 items | published: 88 (87 verified 200, 1 robots-denied and disclosed)