Anthropic opened its internal employee training curriculum to the public on August 20 at academy.claude.com: 355 tutorials spanning Claude.ai, Cowork, Code, Tag, and Platform, per EdTech Innovation Hub. Three anchor courses run from a four-hour AI Fluency track built on Anthropic's 4D model down to a 45-minute beta on human-agent teams, and administrators get org-wide rollout guides. If you are paying for Enterprise seats and building your own onboarding deck from scratch, that deck is now redundant. Point new seats at the curriculum and spend your hours documenting your own use cases instead. See the three anchor courses.

The Salesforce in Claude plugin carries 37 prebuilt sales skills, with meeting prep, deal health, and pipeline review as the flagship three, and it went to public beta this month, per Salesforce. Claude is now the default model behind Agentforce's Atlas Reasoning Engine and behind Slack. Two production numbers landed with the keynote: TSA's Ace agent handles 100,000 traveler conversations a month and resolves 96% of them without human escalation, per the Dreamforce recap on LinkedIn. Salesforce also says its own employees have banked 8.1 million hours of annualized productivity through the Claude-powered Slackbot. Pre-event materials carried almost no customer ROI data, so these are the first named-account figures procurement can cite in a business case. Signups are open to any Salesforce enterprise customer today. See the TSA and Smarsh numbers.

Give a session a budget and it pauses with a budget_reached stop reason rather than issuing more model requests, per Anthropic's release notes for August 7. Change or remove the cap and it resumes, with session state intact, so you can inspect the runaway loop before deciding whether to fund it. The same August run added allowed_domains and blocked_domains to the built-in web_search and web_fetch tools. Cost ceilings and network boundaries are now platform config a compliance reviewer can read in one screen, instead of application logic your developers have to write and maintain. Running agents in production? This is an afternoon of config that removes your two largest unbounded risks. See the two config changes.

Working groups below the CEO level at Anthropic, OpenAI, and Google have met regularly since July 2026 to build a voluntary safety standards body loosely modeled on FINRA, per Seoul Economic Daily and The Information. Three functions are on the table: independent model evaluations, pre-release safety reviews, and risk assessments all three would run before a model ships. A draft executive order for a government-led version got blocked internally by David Sacks, and that is the vacuum this fills. Grade the participants before you grade the plan. The Future of Life Institute's summer 2026 index rated Anthropic C+ at 2.66, OpenAI C at 2.28, and Google DeepMind C at 2.01. Voluntary norms become procurement requirements the moment large buyers put them in an RFP, and that is the milestone to watch, not the launch announcement. See what the body would actually do.

House Speaker Mike Johnson told reporters on September 15 that AI leaders "have got to be responsible on their own side," and asked Trump to convene executives rather than legislate, per ABC News. Trump, in a Fox News interview, called AI safety warnings a "conspiracy." Both landed in a week when senior researchers left Anthropic and Google DeepMind citing safety concerns and multiple tech CEOs were publicly asking Congress to move. Practical read for anyone buying AI: your vendor's published safety policy is the enforcement mechanism for the foreseeable future, so read it before you sign and put the parts you depend on into the contract. Read the full exchange.

Budget caps and domain allowlists solve the "what can the agent spend and reach" half of production Claude. The other half is wiring it into the systems your team already runs, and Make does that with visual scenarios instead of a bespoke integration layer nobody wants to own. Good first build: an agent run that writes its result straight into your CRM and Slack. Try Make →

The platform sending you this email. Clean API, real plain-text support, a web archive that renders on your own domain, and no AMP nonsense to debug. Try Buttondown →

Rated 4.3, one tenth below August, and the drop is governance rather than capability. Solo on Pro: keep it and move Tab off GPT. Small team: stay month-to-month until model access is in writing. Enterprise: Self-hosted Machines clears data residency, so pilot it, then get the guarantee. Read it or try Cursor.