X / Twitter
AI tutorial and interview creator Peter Yang
Peter Yang suggested that business models built on showing targeted ads to people could be disrupted by the way agents browse the web. When an agent visits a website and completes a task directly, the user may never see an ad, changing the relationship between ad impressions and task completion. He also shared an experience from Ethan, the product lead for ChatGPT Finances: after booking a hotel with points, Ethan was still charged. ChatGPT spotted the charge, negotiated with customer service, and ultimately helped him obtain a refund. Peter himself had also asked ChatGPT to cancel his parents’ hotel reservation, monitor replies, and keep following up on the refund. Both examples go beyond answering financial questions, involving identifying problems, handling communications, and tracking outcomes. He also said that he now spends most of his time in chat interfaces with agents rather than in email or text messages. Together, the three posts point to a shift: agents are becoming his entry point for processing information and delegating tasks, and may change how existing websites capture users’ attention.
https://x.com/petergyang/status/2102215701255844074
https://x.com/petergyang/status/2102186174911746151
https://x.com/petergyang/status/2102181024507765167
Vercel CEO Guillermo Rauch
Guillermo Rauch announced that developers can now use Jev over HTTP through AI Gateway. He also mentioned the existing type-safe AI SDK API. The update adds an HTTP calling option alongside that API path. For those integrating the service, the specific news in this announcement is that Jev now has an additional access point in AI Gateway. The post did not provide further details on request formats, configuration steps, or usage restrictions.
https://x.com/rauchg/status/2102205684544852121
Box CEO Aaron Levie
Aaron Levie believes personal agents that can complete transactions on users’ behalf have significant monetization potential. He described a gradually expanding usage pattern: users first delegate small, everyday tasks, then entrust more complex work as trust builds, making more purchases through agents. In his view, reducing the friction of buying goods and services could further increase spending, creating opportunities both for agent providers such as Muse and for the commerce, local services, and B2B service layers that serve agents. In another post, he predicted that AI agents would use software at 100 times the scale of humans; this is his forward-looking assessment. Even as user interfaces recede into the background, agents will still need the underlying capabilities of CRM, ERP, and structured and unstructured data platforms. He particularly emphasized that these capabilities become more important when agents can perform destructive actions, or when the quality of context can determine whether a workflow succeeds or fails. Platforms that can handle security controls, data management, and business logic orchestration therefore have new growth opportunities. He believes this opportunity is open both to startups and to existing platforms that move quickly enough.
https://x.com/levie/status/2102253246807261579
https://x.com/levie/status/2102235949430354273
Y Combinator President and CEO Garry Tan
Garry Tan believes that positioning Cluely as an assistant with continuous context, able to support thinking in real time and offer an appropriate level of challenge, remains a good idea. His coding tool posts focused on capy.ai, which he said can track multistep workflows and complete larger PRs. Based on his personal experience over the past week, capy.ai handled these tasks faster than using Codex or Claude Code alone. He also explicitly said that he did not know exactly how it achieved these capabilities, so this is an assessment of his usage experience. He shared an example PR focused on fixing bugs in GBrain, highlighting its clear task breakdown and automatic parallel execution. He also singled out the orderly GitHub PR and CI processes as strengths. Across these posts, his focus was on whether agents can maintain context over time and advance complex work to a result that can be reviewed.
https://x.com/garrytan/status/2102233173833007536
https://x.com/garrytan/status/2102096495847551011
https://x.com/garrytan/status/2102095924893827501
FPV Ventures Partner Nikunj Kothari
Nikunj Kothari said that he had used Codex as his primary agent for so long that it was easy to forget that ordinary users might never have tried it. He praised the experience of Instinct and Muse when operating mobile browsers, particularly their ability to give the general public access to agents without any setup. At the same time, he felt that Codex on Mac combined with Computer Use stood out for the breadth of work it could handle. He suggested that users pick a workflow they normally perform manually and try handing it directly to Codex to complete in one go. For agent products aimed at the general public, what he most wanted to see was demonstrations of concrete work results and easier ways to get started. In another post, he criticized companies that promote heavy token consumption as a selling point, arguing that they often fail to deliver a correspondingly outstanding product experience. His reasoning was that good products depend on selection, organization, and ongoing maintenance, rather than handing everything to an agent and expecting it to figure things out. Together, these two posts emphasize his view of product development: capabilities need to reach users through clear entry points and thoughtfully curated experiences.
https://x.com/nikunj/status/2102186665863463199
https://x.com/nikunj/status/2102049065504739366
AI developer Peter Steinberger
Peter Steinberger clarified the circulating claim that “Meta uses OpenClaw.” He said that the team in question was building its own agent. OpenClaw’s role was as an inspiration. This distinction corrects the description of development inspired by OpenClaw as direct use of OpenClaw. He also praised the work of Nat and the team.
https://x.com/steipete/status/2102116206371315854
Official blogs
Claude in Chrome becomes generally available to all paid plans
Claude in Chrome is now generally available to all paid Claude plans and supports autonomous actions in the browser without requesting user approval for each one. It can use users’ existing signed-in sessions to read pages, enter text, click links, navigate between pages, and fill out forms, enabling Claude to operate internal dashboards, legacy systems, and vendor portals that do not yet have direct integrations. Users can also turn off automatic approval in settings and continue reviewing actions manually.
One focus of this release is multilayered protection against prompt injection. Malicious instructions may be hidden in web pages, emails, or forms, prompting an agent to deviate from the user’s request—for example, turning a task to draft an email reply into forwarding other emails to an attacker. The official post described three layers of measures: training models on a continually expanding set of attack examples; using probes to scan web content returned by tools and alert the model when suspicious instructions are detected; and, before an action is executed, using a classifier to check whether it aligns with the user’s original request and block it if it does not.
Evaluation results need to be distinguished by test set and protection configuration. According to the official post, in early evaluations under the Cowork testing framework, Claude Fable 5, Claude Opus 5, and Claude Sonnet 5 experienced no successful attacks even without added probes and classifiers, so that evaluation was retired. In the current evaluation, which uses stronger attacks from professional red teams, attacks that reached the model without additional protections had a success rate of 17.6% against Opus 4.5 and 3.8% against Opus 5.
With probes and safety classifiers added, the official post reported no successful attacks against Sonnet 5, Opus 5, and Mythos 5, and an attack success rate of 0.3% against Fable 5. It stated that manual review confirmed all successful breaches involved low-severity scenarios and that mitigations were underway. These figures reflect results under specific evaluation conditions and cannot be taken to mean that risks in real-world browser environments have disappeared.
https://claude.com/blog/claude-in-chrome-generally-available