Top Stories

Everything I Own, Owned

991 points · schlarp.com

A researcher spent a couple of weeks pointing a coding agent at the peripherals sitting on their desk — a microphone, a webcam, a key light — and came away with a plaintext command shell inside the mic, a webcam whose activity LED can be switched off while it records, and a light that hands out memory writes to anyone on the WiFi. The pitch is that peripherals are ideal targets for agentic reverse engineering: they’re tiny computers with a firmware update mechanism, so an agent always has something concrete to iterate against.

HN is fixated on what this signals more broadly — LLM agents are collapsing the cost of firmware RE, which is great for hobbyists auditing their own hardware and unsettling for anyone shipping insecure consumer gadgets. Expect a lot more of this.

Anthropic’s Best AI Model Struggles to Attract Users as Cheaper Tools Thrive

555 points · ft.com

The FT reports that Anthropic’s top-tier model is having trouble winning users against a wave of cheaper alternatives, a sign that raw capability at the frontier isn’t automatically translating into market share when good-enough models cost a fraction as much. It’s the classic price-versus-performance squeeze playing out in real time across the LLM market.

The thread is a good barometer of how developers actually pick models in 2026: latency, cost per token, and API ergonomics often win over benchmark leadership. It also raises the uncomfortable question of how the labs sustain frontier R&D spend when the market keeps commoditizing the tier just below the top.

What Is a Harness?

465 points · earendil.com

If you live in the AI newsfeed, “harness” now means the agent scaffolding around a model — the loops, tools, and guardrails that turn a raw LLM into something that gets work done. This essay steps back to unpack the word itself, from the climbing gear the author wore scrambling up walls in middle school to the software harnesses everyone’s building today, and asks what it really means to control and channel a system’s power.

It resonated on HN because “harness” has quietly become one of the most load-bearing and least-defined terms in the agent world. A clear articulation of what a harness is (and isn’t) is exactly what a lot of people building agents have been missing.

How I Find Problems to Solve as a Staff Engineer

425 points · lalitm.com

A practical field guide to the least-taught part of senior engineering: not solving problems, but finding the right ones. The author lays out how staff-level impact comes from sniffing out the high-leverage work that nobody has framed yet, rather than waiting for tickets to land in your queue.

This is catnip for HN’s large population of senior ICs wrestling with the fuzzy expectations of staff and principal roles. The comments are full of war stories about the gap between “writes great code” and “changes what the team works on.”

A Website for Debloated Open Source Alternatives

351 points · debloat.dev

A curated directory of lean, open-source replacements for bloated mainstream software — the kind of resource that keeps resurfacing because the appetite for lighter, self-hostable, no-telemetry tools never goes away. Think of it as a filter for people who want the functionality without the account requirements, background processes, and ad-tech baggage.

HN loves a good “alternatives to” list, and this one landed at the intersection of two enduring obsessions: open source and fighting software bloat. Expect plenty of comments nominating their own favorite lean tools.

My agent.md to Improve LLM-Assisted Code Quality

324 points · fabiensanglard.net

Fabien Sanglard shares the agent.md file he uses to steer AI coding assistants toward code he actually wants to merge — a concrete, opinionated set of instructions that acts as a persistent style-and-standards guide for the model. It’s a window into how experienced engineers are domesticating LLM assistants rather than just accepting whatever they generate.

The post hits a live nerve: as agent.md, CLAUDE.md, and similar files proliferate, developers are trading tips on what actually moves the needle. Coming from a respected engineer and writer, this one carries extra weight.

How Complex Systems Fail (1998)

328 points · how.complexsystems.fail

Richard Cook’s classic short treatise on why complex systems break — the one that argues failures are rarely due to a single cause and that systems run in a degraded mode most of the time, held together by the constant adaptive work of the people inside them. Nearly three decades on, it keeps getting rediscovered.

It’s back on the front page because its lessons map cleanly onto the AI-infused, deeply interdependent systems engineers build today. The comments are a reliable mix of SRE veterans and newcomers encountering these ideas for the first time.

Andreessen Horowitz Is Investing Billions into a Bleak Future

201 points · modelrepublic.org

A pointed critique arguing that while a16z works to shape US AI policy, its portfolio is stacked with companies that exploit legal loopholes and ship ethically dubious products — from bot farms of fake accounts to tools that normalize cheating on interviews and tests. The piece connects the firm’s policy influence to the kind of future its money is actually funding.

It’s a lightning rod on HN, where a16z’s outsized role in both startups and AI regulation makes it a perennial flashpoint. The thread splits predictably between “this is just how VC works” and “this is exactly the problem.”