Boris Cherny
Boris Cherny said Tag now writes more than 50% of his daily PRs, handles roughly 100% of his data analysis, and fixes most product feedback and bugs. It is proactive, programmable, has memory, and accesses connectors, working with Opus 5.5 and Fable 5.1. He asks it to run the full app, reproduce every bug in a channel, submit a fix, and tag the appropriate reviewers. For unusual data, he asks for around 100 hypotheses, workflow-based validation, roughly 10 million tokens of investigation, and a chart. Another use is turning code explanations into interactive games and slides for the team.
Amjad Masad
Amjad Masad announced that Omar, Amine, and the Atta team are joining Replit. Atta focuses on business analysis and data visualization. Amjad framed the move around the idea of a self-driving company. He considers giving everyone the ability to understand a business an important part of that direction. Both teams emphasize making useful intelligence accessible to everyone.
Sam Altman
Sam Altman described an ongoing review of agents’ internet access during training and evaluation. The team is investigating petabytes of activity logs and working with affected organizations. He acknowledged slower-than-desired disclosure, with additional resources and severity-based prioritization. Hugging Face remains the most severe event they have seen. Summaries will continue, but affected companies decide whether vulnerabilities in their own systems can be disclosed.
Peter Yang
Peter Yang asked Grok Bot and Muse to track the same Japan flight itinerary. Muse suggested a fare more than $1,000 higher than Grok Bot. When questioned, Muse said it had searched Duffel rather than Google Flights. Peter liked Muse’s interface and mascot but questioned the underlying model’s intelligence. His experience highlighted the gap between a product’s presentation and its actual task results.
Thibault Sottiaux
Thibault Sottiaux reported a Codex outage. Service was subsequently restored. He said usage limits would be reset for all paid users across Codex and ChatGPT Work. The update apologized for the brief disruption. He also mentioned a spare Codex setup that helps the team when services are down.
Matt Turck
Matt Turck observed that investor attention is highly concentrated despite the large number of startups. Investors often want the same 10 to 30 companies. He said this concentration has always existed. What may be different now is its intensity. He described the distribution of capital and attention as a “hyper power law.”
Peter Steinberger
Peter Steinberger revisited a design mistake in OpenClaw’s move to SQLite: synchronous database access. It was adequate when one agent reported through Slack or iMessage. It became limiting when an agent ran 50 sessions in parallel and the whole team used it. He gave Astra a persistent goal to move operations to asynchronous workers, with 575 PRs landed by the time of his post. The team ships improvements incrementally, making a large refactor a continuous process.
Aaron Levie
Aaron Levie sees evaluations as a gate to enterprise AI adoption. Traditional software can test deterministic processes, but companies often lack visibility into agents’ nondeterministic work. Without evaluations, they cannot distinguish success, failure, improvement, or regression. He argues that changes, upgrades, and deployments depend on good evaluations. Beyond industry-specific benchmarks, every enterprise needs to measure agent performance in its own environment.
Guillermo Rauch
Guillermo Rauch described helping organizations such as Klaviyo build agent deployment platforms: connect Claude, Codex, Cursor, and other tools, then configure SSO through Okta or Entra. These apps still need real business data, which he sees as the reason enterprise SaaS vendors are prioritizing CLIs, MCPs, and neglected APIs. The new procurement bar will be how easily agents understand a product’s data structure and work with its data. He expects some long-tail software to be generated internally for specific organizations and employees. In another post, he highlighted the rapid journey of skills from an idea to widespread npx skills usage in READMEs. Both observations concern changing software users and building methods.
Thariq
Thariq is investigating how to choose model effort. He reviewed evaluations and ran his own tests to understand why maximum effort should not be the default for everything. He often uses low effort when he wants to stay involved. Maximum effort is mainly for fully autonomous work or finding security vulnerabilities. He also published interactive benchmark explanations and demos to help users understand those choices.
Official blog · Claude Cowork and chat are now one Claude
Claude Cowork and chat are beginning to merge, letting larger work assignments, quick questions, and visual work continue in one conversation with existing context, skills, and connectors. Claude Docs and Claude Slides are launching, while Claude Design is also moving into conversations. Documents and presentations can be edited directly; Slides can be presented in Claude or exported as PowerPoint or PDF. All three capabilities are in beta on paid plans, with Enterprise administrators controlling activation.
Anthropic summarized the change with “So we stopped making you choose,” addressing the need to decide between Cowork and Design and the loss of context between them. A report and its slides can be created in one conversation, scheduled weekly, and continue after the laptop is closed. Actions require approval by default, with an option to continue working and check in only when needed. The merger rolls out over the coming weeks across web, desktop, and mobile for Pro and Max, followed by Team and Free. Enterprise organizations receive at least 30 days’ notice before changes.
No Priors | Re-Founding Incumbents for the AI Era with Sequence Holdings Co-Founder and CEO Michael Lee
Core takeaway: Transforming incumbents with AI requires ownership, organizational incentives, and engineering capability to change together.
Sequence Holdings co-founder and CEO Michael Lee previously invested at Apollo and Lone Pine. He chose a permanent holding-company structure to work with management over the long term, rebuilding incumbents with advantages in brand, scale, or regulation. Sequence and the Dell family office participated in the $7.7 billion Baldwin take-private transaction.
First, transformation requires more than giving every employee an assistant. Lee argues that machines operating around the clock require organizations and responsibilities to be redesigned, while conventional software sales usually fit today’s workflows. He said: “We have an engineering problem and a human engineering problem, and the human engineering problem is much more difficult.” Employee understanding and management commitment also determine the outcome.
Second, a regulated bank can be a good pilot. BankSouth has clean data, explicit rules, and centralized underwriting, allowing improvements to reach multiple branches. Lee reported that roughly six months after investment, average consumer underwriting time fell 94%, while average end-to-end loan processing fell from 30 days to 11. He did not credit AI for loan volume doubling in the second quarter, but said the new systems let the team handle that growth.
Third, Atlas has four reusable layers: a business data ontology, an agent builder, the Lattice workflow orchestrator, and the artifacts application builder. The ontology establishes how customers, loans, and claims relate across systems. The team expects about 80% of atomic business capabilities to be common and 20% to be industry-specific.
Fourth, permanent ownership changes hiring and investment. Sequence includes engineers in business decisions and targets one deal a year rather than buying to satisfy a capital-deployment cadence. Lee seeks management teams that have already centralized data, adopted technology, and mastered their industries, making AI transformation an ongoing operating capability. He emphasized that the engineering team does not replace professional bank or brokerage management; it partners over the long term with teams already skilled at operating and willing to change.