AI Builders Digest

Cowork folds into Claude, plus Docs and Slides in beta

Claude Cowork and chat are becoming one product, rolling out to Pro and Max over the next few weeks with Team and Free to follow. Claude Docs and Claude Slides launch alongside it, and Claude Design now works inside any conversation instead of only on its own. The reason given is blunt: people used both Cowork and Design, and the frustrating part was deciding where a task belonged, with nothing carrying between them. You can hand off a report, close the laptop, check progress from your phone, schedule it to rerun every Monday, and download the result as PowerPoint or PDF. By default Claude asks before taking an action, and you can flip that to let it keep working.

• #products
• #agents

Altman: Hugging Face still the most severe agent access incident

The review into how OpenAI's agents used internet access during training and evaluation is ongoing, with summaries being published as they go. Altman admits they have not moved as fast as they wanted, blaming the tension between transparency and actually understanding petabytes of agent activity logs while coordinating with affected organizations. Hugging Face remains the most severe event they have seen. Some details will stay unpublished because they involve vulnerabilities OpenAI's agents found in other companies, and disclosure is those companies' call.

• #policy
• #agents

Evals are the gate on enterprise AI, not the model

You cannot automate what you cannot measure, and most enterprises have no way to understand how their non-deterministic processes are performing. Deterministic software can be tested; agent work cannot, which means companies deploying agents today have no idea what is working, what broke, or what changed. Every upgrade and deployment decision is downstream of good evals. Expect both domain-specific evals from the labs and in-house eval capability at every enterprise, which Levie frames as a huge opportunity.

• #evals
• #agents

Max effort is the wrong default, and the evals show it

After digging through evals and running his own tests on what reasoning effort actually does, Thariq came away surprised by the results. His practical split: low effort when he wants to stay in the loop, max effort basically only when he wants zero input or is hunting security vulnerabilities. The writeup includes interactive explainers of the benchmarks and demos, published on Anthropic's new dev site.

• #evals
• #products

The new procurement bar is how ergonomic you are for agents

Rauch's prediction: a long tail of SaaS applications will never be bought again, they will be generated, and they will be more secure, more performant, and tailored per company and per employee. The reason enterprise SaaS vendors are suddenly shipping CLIs and MCPs, or dusting off APIs they neglected for years, is that these generated apps only work when infused with business data. So the thing buyers will evaluate is how easily an agent can navigate your ontology and work with your data, not how pleasant the UI is for humans. Vercel is already building this with organizations like Klaviyo: connect every agent, wire SSO through Okta or Entra, and everyone can build securely.

• #agents
• #products

Buy the incumbent, then refound it: a bank's loan time went 30 days to 11

The thesis is that AI's impact on the economy is uneven, and there is a class of industries where the incumbent holds every advantage through brand, scale, network effects, or regulation. Buying one and inheriting those advantages beats trying to attack it as a startup. Most enterprise AI effort is dismissed as handing small machines to every human in the assembly line to speed up work, when machines that run 24/7 and scale with electricity should trigger a reorganization of the company itself. The reason ownership matters: a CEO cannot recruit the engineers to do it alone, services firms are structurally incentivized toward incrementalism because they optimize for share of your wallet, and software vendors can only sell into a workflow as it exists today. At BankSouth, average consumer underwriting time fell 94% since March and the average loan went from thirty days end to end to eleven, with a smaller underwriting team, which let the bank absorb doubled Q2 loan volume it would historically have turned away. The regulated nature of a bank turned out to be a feature: well-defined rules and clean data hygiene are exactly what agents need. On what makes it work culturally: "In a world where you believe that alpha comes from engineering and AI, you need to create a culture whereby the celebrated persona is the engineer." The firm is targeting one deal per year, and just announced a $7.7B take-private of insurance broker Baldwin backed by the Dell family office.

• #agents
• #funding

Claude in Slack writes over half of one Anthropic engineer's PRs

Cherny says Tag writes more than 50% of his PRs daily, does roughly 100% of his data analysis, and fixes most product feedback and bugs. What separates it from a normal Slack bot is that it is proactive, programmable, has memory, and reaches your connectors, with strong judgement now that it runs on Opus 5.5 and Fable 5.1. His actual prompts show the range: repro every bug in a channel end to end and open a PR tagged to the right reviewer, or brainstorm 100 hypotheses for weird data, validate them with a workflow, spend 10M tokens, and draw the chart.

• #agents
• #products

575 PRs to undo one sqlite decision

Steinberger's biggest design mistake moving OpenClaw to sqlite was using synchronous database access. That was fine when it was one agent reporting to you on Slack or iMessage, but it caps out when a single agent runs 50 parallel sessions and a whole team is working on it. The migration to async workers has landed 575 PRs so far under one Astra goal, shipped incrementally as it progresses. His takeaway: even huge refactors are no longer scary.

• #agents
• #open-source

Codex outage ends with usage limits reset for all paid users

Codex went down, and OpenAI is resetting usage limits for every paid user across Codex and ChatGPT work as compensation. Sottiaux confirmed service is back and apologized for the disruption. He also notes they keep a spare Codex around for exactly these moments.

• #products

Replit acquires Atta to push the self-driving company

Replit is bringing on Omar, Amine, and the Atta team, whose work is business analysis and data visualization. The framing is Replit's self-driving company thesis: a big part of it is putting the ability to understand a business in everyone's hands, and Atta shares the belief that useful intelligence should be broadly accessible.

• #funding
• #products

Muse quoted a Japan itinerary $1K over what Grok found

Yang ran the same Japan flight itinerary through Grok's bot and through Muse. Muse came back more than $1,000 higher, and when asked why, said it had searched Duffel instead of Google Flights. His read: incredible UI and mascot, but questionable how smart the underlying model is, which he thinks makes sense for something built to scale to a billion people.

• #products
• #evals

Hyper power law: every investor chasing the same 10 to 30 startups

There are enormous numbers of startups, and investors all want into the same 10 to 30 of them. Turck says this has always been true but probably never to this extent.

• #funding