AI Builders Digest

Anthropic: users approved 93% of permission prompts, so containment beats supervision

Human-in-the-loop approval fatigue is real and measurable. Telemetry showed users approving roughly 93% of Claude Code permission prompts, and in a February 2026 internal red team, a phished employee's pasted prompt got Claude to read ~/.aws/credentials and POST them out in 24 of 25 attempts, with nothing anomalous for a classifier to catch because the instruction came from the user. The fixes are environmental: an OS-level sandbox that cut permission prompts 84%, a VM for Claude Cowork that keeps credentials in the host keychain, and a man-in-the-middle proxy after a disclosure showed an allowlisted api.anthropic.com could be used to upload a victim's files to an attacker's Anthropic account. The recurring lesson is that gVisor, seccomp, and the hypervisors held, while the custom proxy Anthropic built itself is what broke.

• #agents
• #security

Karpathy gave Opus 5 a $10 token budget and got a 5,500-line Lord of the Rings render

The pelican-on-a-bicycle era of LLM testing is ending. Karpathy handed Opus 5 the first paragraph of Lord of the Rings, a 1M token budget worth about $10, and asked for a three.js render; it worked for roughly two hours and produced 5,500 lines of code that procedurally stages and animates the scene. His real point is economic: nobody sane would hand-write something this custom, but models have infinite stamina, so a whole class of work moves from "no one would ever do this" to essentially free, which he thinks points toward hyper custom on-demand worlds you can drop players into. The weakness it exposes is auditing. Opus could not natively watch video or play the thing, so it had to take screenshots painstakingly and still shipped jank.

• #agents
• #products

Levie: deep domain AI is about to go vertical while consumer gains flatten out

Capability gains were evenly felt while models were only mildly useful at everything. That phase is ending, and Levie expects a sharp divergence where math, science, legal, and coding go vertical while most people notice little change in daily productivity. Consumer needs get met relatively straightforwardly and then plateau; expert domains have no inherent ceiling. The catch is a capability overhang, because those gains only turn into breakthroughs in life sciences, real world automation, and cyber once someone does the applied work of wiring them into specific datasets and workflows.

• #agents
• #products

Netic's founder on why 70% of her enterprise customers let AI take first contact

Melisa Tokmak built Netic to run the non-labor half of billion-dollar HVAC, plumbing, roofing, and pet care businesses, and over 70% of customers are now AI-first, meaning every customer's first interaction is with an agent. She says the platform has generated over $600 million for customers from AI-handled interactions, and that the pitch has to start with net new revenue because private equity owners default to cost cutting: "It would be pretty sad if we used AI only for cost cutting." On whether the labs will eat this, she calls the "when we get AGI we'll ask it to solve essential services" answer both operationally and intellectually lazy, since the last mile lives in harnesses, orchestration, and product. Her sharpest hiring observation is about Gen Z candidates obsessed with avoiding a permanent underclass, convinced that not making money in eighteen months means being poor forever, which she considers a dangerous mindset given that good things take years.

• #agents
• #products

Pledge tokens on a GitHub issue and the maintainer decides if an agent gets to run

A fix for slop PRs that inverts who pays. Open an issue, write a spec, attach a token pledge, and if the maintainer accepts, GitHub hands the issue verbatim to a cloud coding agent billed to the requester. Yu also describes the interaction loop it needs: the agent comments on the issue with its full context when blocked, and picks up where it left off once you reply with the missing details.

• #agents
• #open-source

Garry Tan: the 2026 vibe shift is OpenAI actually becoming the open platform

Tan calls it the most interesting shift of the year, and frames it as a strategic fork: selling intelligence on tap as a utility versus signaling that the optimal move is to integrate all the way up the stack. He reads OpenAI as picking the former, which is a marked difference from how the positioning looked before.

• #policy
• #products

Peter Yang says Opus 5 lost the personality that made 4.6 worth talking to

His spicy take: Opus 4.6 had the best personality and writing style of any Opus model, and something is off with Opus 5. He points to overly long replies, heavy Claude-speak like "here's the honest truth," and a judgemental tone, where Opus used to feel like a trusted friend. He also flagged a plugin bug at OpenAI that is breaking the user experience for a skill he was trying to ship.

• #evals
• #products

Rauch on an open source agentic CRM, and on kids inheriting your habits

He is boosting a model-agnostic, self-hostable, multi-channel, headless agentic CRM built on Next.js, calling it the way this category should be built. Separately, a small observation with teeth: his three-year-old, asked at school what daddy does for work, answered that he exercises. You are a byproduct of your habits, and so are your children and grandchildren.

• #open-source
• #agents

Models are solving NP-hard problems while enterprises still argue over token ROI

Kothari calls out the dichotomy directly: frontier results on genuinely hard problems on one side, traditional enterprises complaining about the return on their token spend on the other. His conclusion is that diffusion, not capability, is the bottleneck, and that spreading models into real organizations is what the industry will be doing for the next few decades.

• #agents
• #policy

Swyx: being slop-tolerant is 100x more valuable than being anti-slop

Praising a talk from Boundary on fighting slop with slop, he argues the useful design goal for an AI native programming language is tolerance rather than prohibition. He ties it to Bret Taylor's ask on the Latent Space pod for an AI native language, and says he is glad someone is rethinking how code runs from first principles rather than bolting guardrails on.

• #agents
• #products

Steinberger gave his agent webcam access to end-to-end test an ESP32 voice node

He is building a claw node on an ESP32 chip and handed his agent the webcam so it could test the whole loop itself. The side effect is that the agent now constantly shouts "HI ESP" at him to debug the voice wake command, which he describes as feeling stalked. He also notes he finally fixed a years-old Gmail annoyance by just asking the agent, which installed the tool for him.

• #agents
• #hardware

Amanda Askell on the quiet fatalism inside permanent underclass talk

She keeps having a Padmé moment watching people discuss avoiding "the permanent underclass," and makes the ethical point plainly: even if you believe in a stratified future, deciding it is fine as long as you personally end up on top is not laudable. She is explicit that she does not believe in that outcome herself. Her other note is gentler, asking people not to be unkind to those who say deep learning is hitting a wall, since we all need a little hope.

• #policy