Tool of the Week: Claude's Agent Stack Just Went GA
On August 19, Anthropic moved four pieces of the Claude Developer Platform out of beta at once: computer use, a new browser use tool, the Skills API, and the Files API. None of this is a new model. All of it is the plumbing you'd need to actually run an agent in production instead of a demo.
Computer use lets an agent operate software the way a person would: it sees a screenshot, then clicks, types, and scrolls. That matters for the tools that were never built with an API, which in most small businesses is most of them: the scheduling system, the internal admin panel, the vendor portal nobody ever integrated.
Browser use is a separate, new tool, not a rename of computer use. Instead of just looking at a picture of a web page, it reads the actual structure of the page and acts on a specific field or button. For anything that lives in a browser, that is more reliable than clicking a coordinate and hoping the layout didn't shift.
The Skills API is the one worth sitting with. You upload a procedure once, version it, and pin every request to a specific version_id, or let it run on "latest" while you're still iterating. That is the difference between "an agent that does this task the way I taught it last Tuesday" and "an agent that might have quietly drifted."
The Files API got boring but useful upgrades: a rate limit five times higher, and 1TB of storage.
And the toolset itself got faster. The new version takes several actions per turn (click, type, key, screenshot) instead of one action per round trip to the model. Anthropic says early access customers saw 20-40% fewer round trips per task, meaning the same job finishes faster and costs less to run.
What I'd actually do with this:
If you've been prototyping an agent that clicks through an app with no API, this is your green light to build for production, not just demo it. The reliability story (browser use reading structure, fewer round trips) is specifically what was missing before.
Pin anything that touches money or client data to a version_id. "Latest" is fine while you're experimenting. It is the wrong setting for anything running unattended.
Use browser use for actual web pages and save raw computer use for desktop apps that have no web equivalent. They solve overlapping but different problems now, so pick the one that matches where the task actually lives.
If your agent workflow feels slow, check whether the wait is round trips, not the model. The multi-action toolset is the fix for that specific bottleneck, and it is easy to misdiagnose as "I need a faster model" when the real cost is repeated calls.
Who this is for: anyone building or piloting an agent that has to operate software without a real API: QuickBooks-adjacent tools, internal SaaS, legacy admin panels, client portals. If you had something running on the older beta toolset, budget an hour to test the new one before you assume nothing changed.
Quick Hits
Google just handed its agent-to-agent protocol to the same neutral body that runs MCP. On August 20, Google formally donated its Agent2Agent (A2A) protocol to the Agentic AI Foundation (AAIF), the Linux Foundation-backed body that already hosts Anthropic's Model Context Protocol. A2A governs how agents built by different vendors find each other and hand off work; MCP governs how one agent talks to its own tools and data, a different layer of the stack. AAIF has grown from under 40 members at its December 2025 launch to more than 250, including Google, Microsoft, Amazon, Anthropic, OpenAI, Shopify, and Block. Why it matters: this is standards plumbing, not a new product, but it tells you where the industry expects friction to show up next: agents from different vendors having to talk to each other, not just to their own tools. If anything you build might ever need to hand a task from one agent or vendor to another, both major protocols now sit under one neutral roof instead of one company's roadmap.
A free-forever AI CRM just launched aimed squarely at home-services solopreneurs. BlueFaucet launched August 19 as an autonomous AI agent CRM for businesses under five employees in local services, home services, personal care, and food and beverage. It automates customer retention, marketing, and billing, working as a kind of always-on virtual employee. Free forever for up to 50 customers, no credit card required, 30-day trial beyond that. Why it matters: this is the exact buyer profile a lot of "AI employee" pitches target right now, and it just got a free, packaged competitor. Worth a look both as a competitive signal and as a pricing reference point, whether or not it's a fit for your own stack.
Anthropic opened a free school for using Claude, with 26 courses. Claude Academy launched at academy.claude.com on August 20, replacing the prior Anthropic Academy, with four learning paths running from a first chatbot conversation to a nine-hour track on building with the API. Sign-up needs only an email, and completion badges are shareable on LinkedIn. Why it matters: if you have team members, contractors, or clients who are AI-curious but not yet AI-fluent, this is now a free, structured place to point them instead of building your own onboarding material from scratch.
Prompt of the Week: Screen Automation or Real API?
Computer use and browser use make it tempting to point an agent at any screen and call it automated. Before you do, this prompt checks whether there's a cleaner option, and whether the task is even a good fit for screen-based automation in the first place.
I want to know whether a repetitive task I'm doing in [app name] should be
automated with a screen-based AI agent (computer use or browser use), or
whether there's a real API, export, or integration I'm missing.
Here's the task:
[describe what you do manually, e.g. "every Friday I log into our scheduling
software, pull the week's completed jobs, and copy them into a spreadsheet
for invoicing"]
Ask me anything you need, then tell me:
1. Does [app name] likely have a documented API, webhook, or CSV export that
would do this more reliably than screen automation? What should I search
for to check?
2. If there's no API, is this task stable enough (same buttons, same layout,
week after week) to be a good fit for a screen-based agent, or does the
interface change often enough that it will keep breaking?
3. What's the real risk if this task fails silently one week: is it
reversible, or does it touch money or client data in a way that needs a
human checking the output every time?
4. Based on 1-3: real API integration, screen-based agent, or keep doing it
by hand for now. Be blunt.Run it on the task you're most tempted to automate this week. Half the time the answer is "there was an API the whole time and nobody checked."
One last thing
Like what you're reading? Forward it to one person who'd get something out of it.
And if you want a second set of eyes on where AI could actually save you time, book a free 15-minute audit.