What the people actually building AI said today. One page — a 2-min read.
Claude Duet brought proactive agents closer to everyday work, while builders shared a clearer picture of where those agents help and where they still fall short. AI driven security risks and defenses were the other urgent theme.
Claude Duet opens a US waitlist for a proactive personal agent
Claude Duet is an early research preview for Pro and Max users in the US. Built around Claude Opus 5.5, it uses memory, skills, and connected apps to identify work that needs attention, though Claude says it can still misjudge urgency. Important decisions stay with the user.
Vercel deploys firewall protection ahead of a Node.js patch
Vercel says it rolled out firewall mitigations for a Node.js remote code execution vulnerability before the official patch, automatically protecting traffic to all deployments. It is also monitoring the issue and rolling out managed runtime upgrades. Separately, Vercel added DeepSeek v4.1 Flash to Token Pass for inexpensive subagent work.
Two weeks with Duet reveal a useful agent with clear limits
After two weeks testing Claude Duet, Shipper says he opens it daily and now splits his time roughly evenly between Claude and ChatGPT. He likes its separate app, work channels, task list, and document editing. It cannot direct his desktop Claude sessions, its email drafts lag behind dot's, and voice works only as transcription.
Levie expects AI agents to expand cybersecurity work
Levie argues that agents will help both attackers and defenders inspect code, logs, and traffic at a scale previously out of reach. He predicts enterprises will invest heavily in security systems and people who can direct agents, review findings, patch systems, and respond to incidents.
Every's frontier team turns AI experiments into usable workflows
Every's dedicated frontier team tests new models, measures performance on real work, and turns promising experiments into tools the wider company can use. Its members describe a practical division of labor between rapid experimenters and engineers who make projects reliable. As one member puts it: “If you do a hundred experiments and don't let it land and don't integrate it, it doesn't work.”
Steinberger says GPT-5.6 Sol found a PHP vulnerability
Steinberger says GPT-5.6 Sol found the PHP vulnerability expected to cause trouble this week. His post gives no technical details, but it is a concrete example of an AI model being used to discover a security flaw.
Pigment and Legora announce large Series E rounds
Turck reports a $250 million Series E for Pigment and a $550 million Series E for Legora. An investor in Pigment since its seed round, he argues that tokens, APIs, and agents are pushing companies toward continuous financial planning.
AI may make building more compelling than using
Agarwal watched his son spend about three hours watching Star Wars and about 20 hours making games about it, then barely play them. He thinks AI could make the enjoyment programmers get from creating software available to far more people.
One email a day. Unsubscribe in one click.
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.
The practical work of making agents dependable led today: longer caching, durable runs, and better ways to supervise coding agents. Builders also pointed to on-device voice AI and the growing gap between impressive demos and tools reliable enough for production.
OpenAI extends caching and fixes usage limits for 40,000 accounts
The API now supports 12-hour caching. Logged-in Codex and Work users can return to old conversations and still hit the cache. OpenAI also corrected a plan-limit mismatch affecting about 40,000 accounts and says affected users will receive credits.
Vercel is opening its durable agent infrastructure to other harnesses
Guillermo Rauch says the infrastructure behind Eve and Workflows will support harnesses such as Pi and OpenCode, whether run locally, on Vercel, or independently. It can also make Claude Code and Codex runs durable and observable, though he says embedding those tools requires an extra sandbox and is less efficient.
Aaron Levie sees task-specific agents becoming an enterprise standard
Levie expects autonomous systems dedicated to particular jobs or domains to become a lasting form of AI software. Better tool use, stronger domain judgment, and infrastructure for long-running work now let agents handle more sophisticated workflows, with human oversight still needed.
Gradium launches small, open-source voice models for on-device use
Matt Turck highlighted Gradium's new suite of small text-to-speech models designed to run on devices; he describes them as state of the art but gives no benchmark figures. He also said FirstMark led Daytona's Series B after leading its Series A earlier this year.
Peter Steinberger favors worktrees over subagents for coding work
Steinberger says his strongest workflow had Astra plan and run Codex CLI in worktrees while Sol 6.1 wrote code. A separate agent challenged and monitored the work, an approach he packaged as an OpenClaw plugin called “smart goal.”
Swyx says writing-model restrictions block a production workflow
Swyx says Opus writes well until a prompt contains restricted words, making it unreliable in the critical path of the AI media organization he wants to build. He also tried using Codex to make animated explainer videos and judged the result unsuccessful.
GStack now works with Capy's Captains software factory
Garry Tan says GStack is now Capy-aware and works with Captains, Capy's software factory feature. He offered no further technical details in the post.
Today’s strongest agent stories focused on how teams actually use them: Every launched a shared company agent, while Vercel released tooling to keep agents running through failures. Builders also argued that better models may make specialized software more valuable, and that doing more with agents is not always the same as doing the right work.
Every launches a shared company agent after testing personal agents
Every released the Every Agent to subscribers after finding that individual agents were hard to maintain and scattered useful knowledge across the company. Its team decided to pool that effort: “why don't we take all that energy and put it into one agent that the whole company benefits from?” They expect personal agents to remain useful outside work, while a shared agent learns a company’s workflows.
Vercel open sources its HTML artifact builder
Guillermo Rauch says Vercel has open sourced a fast HTML artifact builder powered by libfx. It packages work the team did to keep v0’s agent running through server crashes, unreliable model responses, and function limits.
Aaron Levie expects better models to increase demand for specialized AI
Levie argues that stronger models make domain-specific AI systems more valuable, because long-running agents need specialized data, tools, routing, and changes to real workflows. He expects autonomous systems to become practical across fields including healthcare, finance, and manufacturing, though products close to model labs’ core uses will face tougher competition.
Wabi 2.0 pairs a main agent with temporary, shareable chats
Nikunj Kothari says Wabi 2.0 clicked for him because its main agent creates chats for specific tasks that can be shared and archived when the task ends. It created a chat around a trip after seeing his flight, then archived it afterward. He likes the structure but says the product is currently too proactive and its apps need work.
Swyx questions whether people really want to move beyond chat
Swyx suggests that calls to end chat interfaces say less about what people will use than about their appetite for something new. His point is that public enthusiasm for a new format may not match the behavior people return to.
