What the people actually building AI said today. One page — a 3-min read.
Claude moved into personal finance, while Box and Replit focused on giving AI systems safer access to data and compute. Builders also pushed a practical theme: useful agents need reliable decisions, the right context, and work they can carry forward over time.
Claude adds read-only personal finance tools in US beta
Claude's new Personal Finance feature lets users connect selected accounts and ask questions about their balances and transactions. Access is read-only, with no ability to move or invest money. The ad-free beta is rolling out gradually in the US on Pro, Max, and Ultra plans.
Box and NVIDIA partner on context retrieval for enterprise agents
Box is partnering with NVIDIA to combine keyword and meaning-based search so agents can find relevant information across enterprise content. Aaron Levie argues that access to documents, contracts, and other unstructured data will determine whether agents can do useful work. He says retrieval will follow existing security, access, and governance rules.
Replit says its custom kernel avoids 96% of Linux zero-days
Amjad Masad says Replit's custom kernel build lets it avoid 96% of Linux zero-day vulnerabilities. He points to a new sandbox design, putting infrastructure security at the center of its AI coding system.
JEV creator says model reliability matters more than benchmark wins
The JEV creator argues that developers need models they can rely on for small decisions inside software, rather than more impressive demos that fail in production. “I think people should be complaining to us so much more,” he says, urging developers to report where models break so builders can improve them. He estimates coding agents offer a 20% to 40% speedup on average, with wide variation by task and user.
Vercel pitches runtime schedules for agents that act over time
Guillermo Rauch calls “doing stuff for you over time” the killer app for agents. Vercel Schedules can be defined at runtime and sit outside the sandbox, letting an agent invoke compute when a reminder or follow-up is due. He also says Grok and Cursor can now build and deploy through a SpaceXSI and Vercel integration.
Sottiaux says formal proofs at massive scale are now worth considering
Thibault Sottiaux argues that formally proving tens of millions of lines of code is no longer a ridiculous design goal. He expects much more secure infrastructure eventually, while warning of turbulence along the way.
Garry Tan urges Congress to open distribution to AI startups
Garry Tan sees a bipartisan chance to curb Big Tech's market power and make it easier for new AI products to reach users. His argument is that advances in AI will spread more widely if thousands of startups can compete on distribution.
Thariq gives a concrete use case for persistent personal agents
Thariq's example is a laundry reminder that tracks when a task was last done, waits until the user gets home on a weekday, and follows up until they confirm or delay it. It shows the value of an agent that remembers state and acts at the right time.
Aditya Agarwal warns founders against accepting any offered valuation
An investor's willingness to fund a company at a particular valuation does not make it the right valuation for that company, Agarwal says. Founders should treat the offer as a choice to assess, rather than proof of what their business is worth.
Source data comes from the open-source project follow-builders by zarazhangrui, released under the MIT license. Summaries are generated by an LLM from that project's public feeds, and the summarization prompts are adapted from it. Every item above links to its original source.
Summaries generated automatically. Read the original before relying on any claim.
