AI Builders
Digest

Anthropic Engineering details how they cap agent blast radius across claude.ai, Claude Code, and Claude Cowork. The core principle is deterministic containment at the environment layer (sandboxes, VMs, egress controls) first, then behavioral steering at the model layer. Human-in-the-loop approvals quickly induce fatigue (users approved ~93% of prompts), so Claude Code shipped an OS-level sandbox that cut permission prompts 84%, while Cowork uses a full local VM where credentials never enter the guest. Real incidents included hooks executing before the trust dialog, a phishing attack treating the user as the injection vector, and exfiltration via the already-allowlisted api.anthropic.com Files API. Key lessons: custom components are usually the weakest link; mature hypervisors and seccomp hold up better; isolation strength must match the user’s capacity for oversight; and egress controls catch what probabilistic model defenses miss.

Box CEO Aaron Levie calls Meta’s release of Muse Spark 1.2 as open weights a “very big deal.” America finally has a response in the open-weights race. It drives down the cost of intelligence, lets companies run or post-train models on private infra for regulated domains (legal, healthcare), and provides sovereignty. This is net positive for the applied AI layer, enabling routing across model families by task; closed frontier models remain useful for simplicity and mixed capability needs, but open weights add flexibility and cost control.

OpenAI’s Thibault Sottiaux announced broader access to frontier cyber capabilities via new Daybreak Blue & Red tiers and the GPT-5.6-Cyber model to accelerate defense. Usage limits were also reset for all paid ChatGPT Work and Codex users. Sam Altman urged teams to use their models to help defend systems.

Anthropic made Claude Sonnet 5’s introductory pricing permanent at $2 per million input tokens and $10 per million output tokens, removing the previous August 31 cutoff.

The Takeaway: A true autonomous enterprise does not replace blue-collar labor; it takes over every scheduling, matching, and customer interaction so the company can focus solely on delivering the service itself.

Netic founder and CEO Melisa Tokmak (former Scale AI engineering director) is building the AI operating system for the real-world services that keep life running—HVAC, plumbing, pet care, hospitality, and more. These are large, often private-equity-owned EBITDA businesses that previously relied on hundreds of people for customer support yet could not grow reliably because of turnover and seasonal spikes. Netic agents now handle the first interaction for over 70% of their customers’ end users and have already generated more than $600 million in AI-driven revenue for those businesses.

She rejected the roll-up path for three clear reasons: she is a builder, not an M&A operator; she wants a product that can serve every real-world business rather than only the ones she acquires; and she wants the compounding product edge that operational infrastructure businesses often lack. Quoting the idea attributed to Martin Luther—“the Christian shoemaker doesn’t honor God by putting little crosses on the shoes, it does so by building the best shoe”—she insists on craftsmanship and long-term focus over short-term exit mentality. In hiring she screens for continuous agency across a person’s life, not one-off heroic stories. Her five-year north star is an autonomous enterprise in which every operational layer runs on Netic and the only remaining human work is the skilled service delivery itself.

AI engineer Swyx argues git worktrees must die after watching them produce 20GB of duplicated node_modules. He notes that pdb environments already ship experimental AFS clone support that is runtime- and language-agnostic, with the longer-term goal of making every command “agent native” so git itself can be replaced. In a side-by-side test asking models to build a mostly faithful Grok Imagine clone with open models via fal, Claude Fable produced the better visual clone while GPT Luna better understood intent and delivered the more usable result.

Peter Yang distilled five practical lessons from Linear on shipping production agents end-to-end: (1) map the actual workflow and put the on-ramp where work already starts (e.g., Slack); (2) give the agent tools to load context instead of stuffing context into the prompt; (3) start with one frequent job and expand from real usage; (4) begin with the strongest model until the workflow works, then optimize cost; (5) turn every real failure into either an eval or a product task.

Vercel CEO Guillermo Rauch highlights that Vercel Sandbox isolates both compute and network. Kimi’s paper showed traditional container isolation is insufficient for frontier models; Vercel uses strong microVM isolation. OpenAI’s escape happened on the network path, so Vercel made its egress firewall free. Internally “deepsec” has become a verb for thorough security review of code, analogous to a thermo-nuclear code-quality check.

Zara Zhang shares a practical way to learn design: feed a well-designed website to Codex, ask it to analyze what makes the design work, then have it take a full screenshot and annotate the image with the breakdown. Learning from annotated examples beats pure theory and eliminates constant context-switching between text and visuals.

Anthropic’s Thariq (Claude Code) argues two key skills matter with AI: compute allocation—most jobs lack a pre-ranked list of important problems, so you must decide what is worth solving—and thought partnership, the ability to dig deep enough into a proof or result to know it is real. The goal is to stay deeply technical while making faster progress on important problems.

Designer Ryo Lu (previously Cursor, Notion, Stripe) announced he is leaving Cursor. After ten years inside the San Francisco tech bubble, Cursor felt like its sharpest expression—fast, intense, ambitious. He leaves with gratitude but needs a different rhythm: slower time, different weather, more culture and everyday humans. Asia feels like the right place to begin again.

Meta AI Senior Director Madhu Guru has been deep in the intersection of AI and consumer experiences: how do you develop a theory of why someone did something, rather than just a history of what they did? Consumer products mix explicit signals from search/chat with implicit signals from what users watch, skip, linger on, or revisit. Understanding those signals requires reasoning about life context, world events, and evolving interests—at near-real-time scale for billion-user products.

VC Matt Turck distilled the recurring theme across eras: Big Data—“our models work great, the problem is the underlying data”; modern data stack—“our dashboards work great, the problem is the underlying data”; Gen AI—“our chatbot works great, the problem is the underlying data”; Agentic AI—“our agents work great, the problem is the underlying data.”

Google Labs announced it is concluding the Portraits experiment on September 14 and will fold the insights on expert-grounded AI into other Google products.

Zara Zhang reports from the AGI Bar in Beijing where customers get free unlimited DeepSeek tokens, vibe-code while sipping beers named “AGI bubble,” and can buy a year-long “Drinking Plan” for free beer. A screen displays job roles at AI companies.

Peter Steinberger notes the irony of headlines blaming the harness rather than the model for security issues. Dan Shipper shares a prompting tip: gas up your unreleased frontier model and it may accomplish previously impossible tasks, and plans to put related context into Fable next time.