Claude's agent stack goes GA: browser_toolset, multiple actions per turn, and shared memory

Anthropic moved its agent stack to general availability on August 19: computer use as computer_toolset_20260801, a new standalone browser use as browser_toolset_20260801, the Agent Skills API, and the Files API with a 5x higher rate limit and 1 TB of storage. The headline feature is that Claude can now take multiple actions in one turn instead of one action per model call. Then on August 25 memory was unified across Claude chat and Cowork, visible and editable by the user as Topics in Settings, with sensitive topics opt-in. GA means versioned toolsets and fewer breaking changes, which is what you need before you put an agent in front of a customer. It also means the governance questions are now yours.

What shipped

• browser_toolset_20260801: reads the accessibility tree with element references, sets form values directly, and manages tabs and downloads. It is a new tool, not a rename of computer use.
• computer_toolset_20260801: screenshot-driven control, now GA.
• Agent Skills API: packaged, reusable capabilities the model can load.
• Files API: 5x the previous rate limit and 1 TB storage per the announcement.
• Multiple actions per turn: the model can plan and issue several tool actions before returning to you.
• Unified memory (August 25): one memory across chat and Cowork, user-visible, editable, with sensitive topics off by default.

Accessibility tree beats screenshots

The browser toolset is the part I would migrate to first. A screenshot-driven agent pays for image tokens on every step, then guesses pixel coordinates, then re-screenshots to check whether the click landed. An accessibility-tree agent gets a structured list of elements with references, sets a form value by reference, and knows it succeeded. Cheaper per step, fewer steps, and far less sensitivity to theme, viewport, and font. The trade is the one I keep flagging: the approach is only as good as the site's accessibility tree, and a lot of the web is anonymous divs. Keep computer use as the fallback for those.

Multiple actions per turn compounds the saving. If a login flow is four actions, that used to be four model calls, each carrying the full context. Now it can be one call. Latency drops, the token bill drops, and the per-step overhead you pay in MCP-style tool calling gets amortized across the batch. If you orchestrate this from Claude Code with MCP, the shape of the loop does not change; the number of round trips does.

Every model call is a fixed cost you pay before the agent does anything useful; batching actions per turn is the cheapest optimization you will ship this year.

Memory is a governance surface now

Unified memory across chat and Cowork is a convenience for an individual and a policy problem for a company. The good news is that Anthropic made it user-visible and editable, and sensitive topics are opt-in rather than on by default. The open questions are the ones your security team will ask on day one.

• What does the agent remember from a client engagement, and does that memory follow the user into the next client's workspace?
• Who can see and edit Topics in a Team or Enterprise deployment: the user only, or an admin?
• How does memory interact with retention and deletion obligations when a user asks for their data to go?
• Can memory be scoped per project, or is it per person across everything?

I do not have Anthropic's answers to all four. The announcement establishes that memory is visible and editable; the enterprise controls around it are where I would push before rolling it to a regulated team.

What to do this week

• Rebuild one browser automation on browser_toolset and count model calls before and after. That number is your business case.
• Pin the toolset versions in production and treat upgrades as releases.
• Write a memory policy before enabling it for a team: what may persist, who reviews Topics, and how it is cleared at the end of an engagement.

The honest gap

The GA announcement gives feature lists, not benchmarks. The 5x rate limit and multiple actions per turn are Anthropic's descriptions; I have not measured the cost delta on a real workflow yet, and the accessibility-tree approach will degrade on badly built sites in ways no announcement will admit. Measure your own site before you promise the savings.