Anthropic AI assistant Claude
Claude’s official account says Claude Data, formerly Dashboards, is generally available on paid plans. In addition to dashboards, it can draft data stories and narrated videos from metrics for users to review before sharing. Certified queries and live sharing remain in beta. Interactive connectors are rolling out on web and desktop, letting users open tools from Adobe, Atlassian, and Box beside the conversation and make changes themselves or with Claude’s help. The account lists Whiteboards and Sheets among the updates and says availability may vary while the rollout continues.
Builder Zara Zhang
Zara Zhang uses a 20% versus 80% split to argue that users sometimes fail to return simply because they forgot the product exists. The post gives no data source for the split, which expresses her view on retention. She also suggests sharing work before it is finished so feedback can help decide whether to complete it. Both posts treat giving people another chance to encounter the product or work as part of building it.
AI Engineer organizer Swyx
Swyx announces AI Engineer livestreams from 9 a.m. to 6 p.m. on the following day and on Thursday. He says a Thursday keynote was subject to an embargo. The post does not give its substance or specify a time zone. Separately, he says his agents keep choosing Resend when he wants a different email service and puts that question to Zeno Rocha.
OpenAI Codex product team member Nan Yu
Nan Yu shares an implementation detail about dictation shortcuts. Users can map the Apple keyboard’s dictation key, which he describes as a special variant of F5. The team also supports push-to-talk, double-tap, and single-tap shortcuts. He credits a colleague for getting those keyboard details right.
AI tutorial author Peter Yang
Peter Yang argues that the role AI plays in making a game matters. He distinguishes using it to churn out low-quality work from using it as one tool in creating a good game, and supports the latter. The post offers a view about creative work without assessing any particular game or production process.
Box CEO Aaron Levie
Aaron Levie describes a progression from chat to agents that execute complete tasks and then to groups of agents working toward a goal. He thinks work that can be completed in a self-contained digital environment and is constrained by research or data throughput stands to benefit most. He sees opportunities in harnesses, applications and implementation work. His claims about output ceilings and industry reach are forecasts. He separately announces a Box partnership with Anthropic that lets Claude view Box content and write file updates back through MCP. He cites contracts, presentations, and strategy documents as examples and expects agents to interact with enterprise files far more often. The integration provides read and write access subject to the permissions granted.
Designer Ryo Lu, formerly of Cursor, Notion, and Stripe
Ryo Lu argues that software can help people participate in creative work and attempt things previously beyond their reach, as well as complete tasks. He draws an analogy with filmmaking and music. A separate post emphasizes collaboration among people and worries that everyone managing their own agents could leave work siloed. His focus is whether people can create together and whether tools broaden participation.
Vercel CEO Guillermo Rauch
Guillermo Rauch reports that Vercel Security Review identified 57,897 vulnerabilities in less than a week after launch, classifying 6,029 as critical or high and saying agents had patched 14,665. He says it can be enabled on any repository regardless of deployment platform. The figures come from the vendor, and the post does not give false-positive rates or patch verification methods. He separately praises the Rust ecosystem, which predates agentic coding. He expects agents to expand the use of purpose-built languages and make direct work in C, LLVM IR, or assembly more practical in some cases.
Every CEO Dan Shipper
Dan Shipper reports Every’s first-day tests of Gemini 4 Carbon. In one Senior Engineer benchmark run, Carbon scored 44% and Argon 21%. Given a backlog in another trial, it marked six tasks ready and completed 15, including work already assigned elsewhere. The team stopped it after nearly four and a half hours. It also built a browser-based model from 3D scans and an equipment list to check whether new gear would appear in shot. The team liked some of its large-project tests and documentation, but found that it sometimes produced very large code files. Audio and video results showed promise, interface design drew mixed reviews, and the writing benchmark put it near Argon. Shipper plans to use it occasionally on large jobs with a clear finish line and review the results.
Google VP Josh Woodward
Josh Woodward says Gemini 4 Carbon was available at the time to Ultra subscribers worldwide in the Gemini App. His announcement covers availability in the app. The post does not give a timeline for API access or other plans.
Former Claude Code contributor Thariq
Thariq wants more tools that show how people and models can build something together. A one-shot demo can leave that collaboration offscreen. His focus is whether the tool helps someone stay involved and progressively realize their own idea.
Latent Space — Designing for Agents — Zeno Rocha, Resend
Key takeaway: Zeno Rocha treats agent experience as an end-to-end design problem, from reading prices and authorizing access to observing what happens after sending mail.
Resend co-founder and CEO Zeno Rocha worked in front-end development and created the Dracula editor theme. He says that while building the sending API, the team found that email templates were the first thing holding developers up. They built the open-source React Email project and documented how to send through providers including SendGrid and Mailgun. Developers could adopt the templates in an existing project and encounter Resend later when they needed a different sending service.
Making a page legible to an agent starts with checking what data the agent actually receives. Rocha asked a model to compare the cost of sending 1.2 million emails, and it misread pricing behind a JavaScript slider. He added a Markdown table to the pricing page and got the right numbers when he asked again a few days later. This single trial exposed a legibility issue. He also notes the maintenance cost of keeping the web page and Markdown in sync.
A remote MCP integration required the team to revisit OAuth and access to individual resources. The server grew from about 20 tools or fewer to more than 80, adding dashboard actions such as checking usage, managing billing, and inviting teammates. Rocha opens tool calls to inspect their output and found a response too large for the model’s context. Coverage and output volume need separate checks. He says, “I want my product to be formless.”
Sending also involves mailbox-provider throttling and domain reputation. Rocha gives the example of sending 100,000 messages to Gmail at once triggering throttling, so batches and queues must account for each provider. When customers share a click-tracking domain, one abuser can affect the rest. Resend supports custom tracking domains to address this. The team is also building more of the underlying sending stack to gain finer performance and delivery data. The interface lets an agent send. Queues, permissions, and delivery feedback determine how much of the process a user can see and control.