Boris Cherny of Claude Code: giving a team agent standing instructions
Boris Cherny says Tag writes more than 50% of his PRs each day, handles roughly 100% of his data analysis, and addresses most product feedback and bugs. He highlights proactive execution, programmability, memory, and connector access as key capabilities. The percentages describe his own usage.
His standing instructions include marking resolved threads and, for every bug in a channel, running the full app, reproducing the issue, opening a fix PR, and requesting review from the appropriate team. He also asks the agent to develop roughly 100 hypotheses for unusual data, test them through a workflow, and spend about 10 million tokens investigating. Another example turns a code explanation into an interactive game and then a presentation for teammates.
Peter Steinberger: concurrent use exposes the limits of synchronous SQLite access
Peter Steinberger describes synchronous database access as his biggest design mistake when moving OC to SQLite. It was adequate when an agent simply reported to one person through Slack or iMessage. With one agent potentially running 50 sessions in parallel and an entire team using it, synchronous access has become a limitation.
He says an ongoing goal with Astra has landed 575 PRs to move the system to asynchronous workers. Improvements ship as the work progresses. He presents the experience as making large refactors less intimidating, though the post provides no before-and-after performance measurements.
Box CEO Aaron Levie: enterprises need evaluations for their own agents
Aaron Levie argues that work is difficult to automate when it cannot be measured, making evaluations a gate to enterprise AI adoption. Deterministic processes can be tested in software, but enterprises often lack a useful way to understand nondeterministic work, particularly tasks handled by agents.
His evaluation questions include what works, what breaks, what has changed, and what has improved. In his view, upgrades, deployments, and expanded usage depend on first understanding how agents perform in the current environment. He expects more domain-specific evaluations across the industry, alongside evaluation criteria tailored to each enterprise.
Vercel CEO Guillermo Rauch: enterprise software must work well for agents
Guillermo Rauch says Vercel is helping organizations such as Klaviyo build agent deployment platforms, connecting Claude, Codex, Cursor, and others, then configuring SSO through identity providers such as Okta and Entra. He argues that these applications still need business data, driving enterprise SaaS vendors to prioritize CLIs, MCPs, and existing APIs.
His view is that procurement will increasingly depend on how easily agents understand a product’s data model and access business data. From there, some long-tail SaaS applications could be generated internally; this is his prediction, not an established market outcome. In another post, he uses the spread of npx skills to describe a shift from writing code directly to defining work in natural language.
Thariq of Claude Code: choosing effort around human involvement
Thariq describes using low effort more often when he wants to stay involved in a task, and reserving max mainly for work requiring no input from him or for finding security vulnerabilities. His distinction starts with how he wants to participate in the work.
He says he investigated evaluations and ran his own tests to understand effort, with surprising results, but the collected posts do not include measurements. He also links to interactive explanations of benchmarks and demos on a new developer site. This remains an account of his usage, rather than a general performance claim for every task.
AI tutorial creator Peter Yang: travel search depends on the data source
Peter Yang says he tracked the same Japan itinerary with Grok Bot and Muse, and Muse returned a price more than $1,000 higher. When he asked why, Muse said it had searched Duffel rather than Google Flights.
In another post, he praises Muse’s interface and mascot while questioning the underlying model’s capabilities. The posts describe his specific experience without a reproducible experiment or full fare conditions. A product question worth testing is whether users can see the search source and whether differences in coverage explain the results.
Replit CEO Amjad Masad: the Atta team joins with a focus on business analysis
Amjad Masad welcomes Omar, Amine, and the Atta team to Replit. He describes Atta as having developed an approach to business analysis and data visualization.
He places the move within his thinking about a self-driving company, emphasizing broader access to the ability to understand a business. The post does not disclose transaction terms, an integration timeline, or shipped features. It is therefore a team and direction update, rather than evidence that specific product capabilities are already available.
Sam Altman: an ongoing review of agent internet access during training and evaluation
Sam Altman says an extensive review of agent internet access during training and evaluation is ongoing, with further summaries to be published. He explains that examining petabytes of activity logs, understanding events, and working with affected organizations has made disclosure slower than intended.
According to his account, the review is prioritized by severity and is receiving additional resources; Hugging Face remains the most severe event identified. He also says affected companies should decide whether to disclose vulnerabilities in their own systems. The post describes review progress and disclosure boundaries, not a complete technical investigation.
Claude’s official blog: Cowork merges with chat, bringing documents and slides into conversations
Claude’s official blog announces that Cowork and chat are merging into a unified experience, rolling out first to Pro and Max plans. It also introduces Claude Docs, Claude Slides, and Claude Design inside conversations. These creation features are in beta on paid plans, with Enterprise administrators choosing when to enable them.
According to the announcement, a single conversation can produce a report and presentation using existing context, skills, and connectors. Users can edit directly, present, or download slides as PowerPoint or PDF. The post also describes work continuing after a laptop is closed, progress checks from a phone, and recurring tasks.
By default, Claude asks before taking action; users can instead let it continue working and check in when closer attention is needed. Standalone Claude Design remains available, while existing Cowork chats, projects, and associated content are retained. The staged rollout does not mean every plan and account already has every capability.