Claude's Agent Toolkit Goes GA, Perplexity Rebuilds Itself as an Agent Platform, and Anthropic Bets on Trust Infrastructure

Three stories from this past week don't share a model release or a lawsuit, but they share a direction: the companies behind the major chatbots are increasingly building for the developers standing on top of them, not just the people typing into a chat box. One is a toolkit graduating from beta, one is a search company re-describing itself as an agent platform, and one is a lab spending real money on the question of whether any of this is good for the people using it.

1. Claude's Agent Stack Leaves Beta

Anthropic moved computer use, the Skills API, and the Files API to general availability on the Claude Platform this week, and added a new browser use tool alongside them. Computer use lets Claude operate an application the way a person would — clicking, typing, and reading the screen — for the large share of enterprise software that has no API at all. Browser use is a separate tool built specifically for the web: instead of clicking at screen coordinates, it reads a page's structure and acts on a named field or button, which holds up better across redesigns and dynamic layouts. The Skills API lets an agent load a folder of instructions, scripts, and templates only when a task actually calls for it, rather than stuffing everything into the system prompt up front. The Files API's GA release adds automatic file expiration, five times higher rate limits, and 1 TB of storage per organization.

The headline number is round-trip count: early-access customers saw 20-40% fewer round trips per task, because Claude can now chain several actions — click, type, read the result, click again — in a single turn instead of one action per API call. That's a real cost and latency change for anything doing UI automation today.

// Before GA: one action, one round trip
await agent.act({ action: 'click', target: '#submit' });
const state = await agent.observe();
await agent.act({ action: 'type', target: '#search', text: query });

// After GA: chained actions in a single turn
await agent.act([
  { action: 'click', target: '#submit' },
  { action: 'type', target: '#search', text: query },
  { action: 'screenshot' },
]);

If your product automates a no-API tool today by scripting individual clicks, this is worth a fresh look — the same task may now cost meaningfully less and finish faster with no change to what it does.

2. Perplexity Stops Calling Itself Just a Search Chatbot

Perplexity spent the week repositioning its developer surface as a "full-stack, model-agnostic platform for building AI agents," organized around four pieces: an Agent API for orchestrating multi-step workflows, a Search API for real-time web-grounded retrieval, an Embeddings API for retrieval at scale, and a Sandbox API for secure code execution, still coming soon. Alongside that, it expanded model access inside the product to include GPT-5.6 Terra and Luna, Grok 4.6, DeepSeek V4 Pro, Kimi K3, and NVIDIA's Nemotron 3.5 Lightning, and shipped a feature called Computer in Email that lets a user start or continue an agentic browsing session directly from an email thread.

The strategic move here isn't any single API — it's the shift from "a chatbot with a good answer engine" to "infrastructure other people build agents on top of," which is the same move Anthropic and OpenAI have already made with their own platform products. For builders, it means the model-agnostic layer you may already be maintaining to swap between labs can now also route retrieval and agent-orchestration calls, not just chat completions — worth checking whether Perplexity's API belongs in that routing table alongside the frontier labs.

3. A Lab Spends on Whether Any of This Is Good for Users

Separately, and less about capability than about what capability is doing to people, Anthropic launched a $5 million grant program funding independent research into how AI affects user wellbeing — giving grantees direct funding, model access, and technical support to build open-source evaluations, rather than relying only on internal metrics. In the same week it launched Claude Academy, a free hub of courses and badges aimed at teaching people to use AI well rather than assuming competence, and rolled out editable memory topics with a dedicated sensitive-topics setting across chat and Cowork.

None of these three moves individually change what your product can do today. Taken together, they're a signal that "does this help or hurt the person using it" is becoming something labs fund and measure rather than assume — and if your own product handles emotionally sensitive conversations, an externally-built, open-source wellbeing eval is a cheaper starting point than building one from scratch.

What This Means for Builders

None of this week's stories require a rebuild, but each is worth a direct check against what you're running. If you automate UI workflows through computer use today, re-benchmark round-trip count and cost against the new GA pricing before assuming last month's numbers still hold. If you maintain a model-agnostic routing layer, decide whether Perplexity's Agent and Search APIs belong in it alongside the frontier labs, not just as a fallback search provider. And if your chatbot handles anything emotionally sensitive, look at what the new wellbeing grants produce before building your own evaluation harness from zero.

— Maya