Editor’s guide and background reading
This edition examines Claude Duet’s early product design, changes in security as agents gain capability, and small frontier teams supporting enterprise AI adoption. Start here for personal-agent boundaries and organizational practice; product experiences and usage forecasts remain attributed to their original authors.
Ruan Yifeng on independent judgment and verification (Chinese background note): https://signaltobuild.com/knowledge/kb-688979357ba2cea0a522eb10870d4da2
Responsibilities and review in AI product teams (Chinese background note): https://signaltobuild.com/knowledge/kb-37fc094ac531db06b0162d108b3aa92b
Source 1 ↗ · Source 2 ↗ · Source 3 ↗ · Source 4 ↗ · Source 5 ↗ · Source 6 ↗
SPC general partner and Bevel Health cofounder Aditya Agarwal
Aditya Agarwal notices that making things with AI may be as enjoyable as using the finished result. His son watched Star Wars EP IV for about 3 hours, then spent about 20 hours making games inspired by it. The child played the games relatively little and was more interested in building them. As a programmer, Agarwal understands the pleasure of programming itself and has felt sorry that many people could not experience it. He asks whether AI might make that creative experience available to everyone.
FirstMark investor and MAD Podcast host Matt Turck
Matt Turck argues that tokens, APIs and agents are pushing financial planning toward a continuous process. He positions Pigment as a platform for that change and praises Eleonore, Romain and their team. He says FirstMark has invested since the seed round. Another post lists announcements of a $250 million Series E for Pigment and a $550 million Series E for Legora. Those amounts come from his announcement roundup; the supplied text adds no revenue, valuation or operating figures.
Official account for Anthropic’s AI assistant Claude
Claude’s official account says Claude Duet is designed around Claude Opus 5.5, which decides what needs the user, what can wait and when to ask first. It carries over memory, skills and connected apps from a Claude account to reduce setup. The company says Duet never directly sees passwords, a new classifier guards against sensitive-information sharing and important decisions remain with the user. These are the company’s descriptions of its safety design, with more details promised. Duet is an early research preview offered through a waitlist to Pro and Max subscribers in the US. The company acknowledges that it can still misjudge urgency and says it will adjust the product using early-user feedback.
Product builder Zara Zhang
Zara Zhang shares the idea of turning a photo into a 3D popup card. In a follow-up she says it was made with a single prompt and links to that prompt. In another brief observation, she calls the personal website the new resume.
Y Combinator CEO Garry Tan
Garry Tan posts two criticisms of Gabriel Zucman’s involvement with California’s Prop. 40. He describes Zucman as an economic adviser to the union sponsoring the measure, a contributor to research behind its revenue estimates and a public campaigner for it. He also points to Zucman’s position at UC Berkeley and argues that reduced state tax revenue would eventually hurt public universities. He cites an aggregate $39 billion net-loss estimate, without a calculation in the supplied posts. His conclusion is that these policy ideas would harm California; the role descriptions, figure and projected consequences are his public claims.
Every CEO Dan Shipper
Dan Shipper says Every has tested Claude Duet for two weeks, during which his use shifted from mostly ChatGPT to roughly an even split between ChatGPT and Claude. He considers Duet a personal agent worthy of being called a professional tool and opens it daily. It has a separate app from Claude on iOS, Android and the web, with channels for areas of work and life rather than a single chat. A todo list exposes task status, and the app supports writing and editing documents, slides and other material. However, he says it is not yet a chief of staff: it can see Claude sessions on his computers but cannot direct them. He also finds its email drafts worse than dot’s and says voice is transcription only, without spoken conversation. Separately, he advocates a frontier team that tracks model changes while the rest of the organization uses what that team learns. He lists experiments with personal benchmarks, explainer videos assembled from Slack and meeting notes, a tabless browser, audience-aware writing tools and a personal agent orchestrator, emphasizing the division between exploration and day-to-day execution.
Box CEO Aaron Levie
Aaron Levie argues that cybersecurity will be one of AI’s most consequential applications. Agents can continuously hunt for, discover and patch vulnerabilities, while attackers can use the same capabilities to find gaps faster. Defenders can analyze logs, inspect code line by line and review traffic to detect anomalies at a larger data scale. He predicts that this loop could consume 100 times more tokens than people currently imagine; that is a forecast rather than measured usage. He expects enterprises to place greater importance on networks, systems, software, data, identity and devices. In another post, he says increasingly autonomous work will still require more talent to prioritize tasks, review findings, patch systems and respond to incidents. He therefore sees opportunities both for systems that protect enterprises and data and for professionals who deploy and operate those defenses.
Practical AI tutorial and interview creator Peter Yang
Peter Yang says he is using AI to secure all his accounts before something goes wrong. He also says he finally paid for 1Password. The post does not give configuration steps or verification results.
OpenClaw and OpenAI builder Peter Steinberger
Peter Steinberger says the PHP vulnerability expected to cause serious disruption that week was found with GPT-5.6 Sol. The post does not describe the vulnerability or discovery process.
Vercel CEO Guillermo Rauch
Guillermo Rauch announces that Token Pass now includes DeepSeek v4.1 Flash (1017), describing it as fast, inexpensive and served with zero-data retention, or ZDR. He recommends using larger, more expensive models to devise plans and this model to run subagents that execute them, illustrating the idea with repeated TypeScript-to-Rust ports. On security, he says the industry coordinated through Slack weeks in advance to protect infrastructure and mitigate php2shell. He reports that Vercel WAF has blocked attempts against PHP origins and urges teams to practice patching and redeploying. In a separate Node.js RCE announcement, he says Vercel rolled out firewall-level mitigations before the official patch release. He says traffic for all deployments is automatically protected while the team monitors and rolls out managed-runtime upgrades. The two vulnerabilities, mitigations and runtime upgrades concern distinct work; the posts do not claim that a firewall completes every software patch.
Anthropic Claude Code and Cowork builder Cat Wu
Cat Wu says the team has heard users’ requests for Claude to take more work off their plate proactively. She describes Claude Duet as a large step toward noticing work and doing it before the user asks. Her post includes a waitlist link.
Google AI tools and experiments team Google Labs
Google Labs releases Dreambeans Roulette: Part 2, continuing the outdoor experience with Sahana, Brian and Nala. This installment explores how Dreambeans stories are made and how to improve them. It also raises a product-design question: can apps help people get off their phones? Part 3 will feature Brian’s field notes and more about the experience and vision behind Google Labs. Dreambeans is currently available in the US to users aged 18 and over, with more regions planned.
AI & I by Every — How a Frontier Team Can Help Your Company Keep Up With AI
Core takeaway: give a small frontier team responsibility for exploring new models, then turn useful findings into tools the rest of the company can adopt. Every CEO Dan Shipper and enterprise AI adoption lead Natalia Quintero describe the roles and concrete outputs of a team that has been operating for several months.
First, explorers and architects complement one another. Dan calls fast experimenters pirates and the people who turn messy prototypes into understandable, maintainable systems architects. Many experiments can be discarded; useful ones move into products and services without requiring every colleague to track each model release.
Second, evaluation should return to real work. Head of evals Mike Taylor turns qualitative vibe checks into personal benchmarks. He used Opus to search Slack and meeting notes, build a source library and story spine, then assemble a Chex explainer video. Each sentence could be traced to notes or meetings, and screenshots came from existing material. He also says that producing more makes protecting colleagues’ attention part of his job.
Third, operations turns experimentation into organizational knowledge. Frontier Wrangler Arielle Shipper gathers token usage and costs across harnesses, broken down by person, model, effort and work category. Individuals approve the classifications before the team assesses whether the spending was worthwhile.
Fourth, tools should adapt to actual users. Senior editor Jack Cheng built a tabless browser with a temporary peek mode to reduce switching. Katie Parrott adds audience.md to Compound Writing’s voice guide and style.md, checking whether drafts serve both early learners and operators. Jannik Jung’s Minga converts goals into dependency-aware plans, works in the background and asks for feedback when a decision is needed.
Fifth, experiments must be adopted. Kieran Klaassen moves from collaborative editors into video, meeting-recording and music tools while continually looking for model failure boundaries. He says, “If you do a hundred experiments and don’t let it land and don’t integrate it, it doesn’t work.” A frontier team therefore also needs to listen to colleagues, explain its findings and bring a few useful results into everyday work.