Confirmed
Google's AI-search opt-out doesn't clearly cover Top Stories that surface inside AI Overviews
Google is rolling out a Search Console setting that lets site owners exclude their content from Search generative AI features. The control removes a site from AI Overviews, AI Mode, and generative Discover results without affecting traditional ranking. Google has not confirmed whether opting out also removes a site's Top Stories carousel when that carousel renders inside AI Overviews, which Search Engine Journal says now happens on roughly one in six U.S. trending-news queries.
In plain termsGoogle now lets a site say 'don't use my pages in AI answers,' but it hasn't said whether that also hides the site from the separate Top Stories box when that box shows up inside an AI answer instead of on its own.
Why it mattersNews and trending-topic publishers weighing this opt-out need Google to answer the Top-Stories-inside-AI-Overviews question before they can judge the real cost of leaving.
Our takeThe open question is whether the opt-out is scoped to generative features only or captures anything that renders inside an AI Overview container, and that distinction now matters for a meaningful share of news queries, not an edge case. Publishers should test the setting against their own Top Stories placements before assuming either answer, since Google's own documentation stops short of saying which way it resolves.
Confirmed
DeepSeek ships V4-Flash-0731, a 304B open-weight model that beats a larger rival on agent benchmarks at a fraction of the price
DeepSeek released DeepSeek-V4-Flash-0731 on July 31, 2026, a 304-billion-parameter open-weight model under an MIT license, priced at $0.14 per million input tokens and $0.27 per million output tokens. Artificial Analysis benchmarks rank it above the larger 428-billion-parameter MiniMax M3, and it scores 82.7 on Terminal-Bench 2.1, a coding-agent benchmark, and 70.3 on Toolathlon-Verified, a tool-use benchmark; the model card attributes the gains to substantially enhanced agentic capabilities.
In plain termsParameters roughly measure how large and capable a model is; this one is large but still costs a fraction of a cent per thousand words to run, and it is tuned specifically for tasks where the model has to use tools or take multi-step actions, not just answer a question.
Why it mattersA model this cheap clearing agent-specific benchmarks ahead of a larger rival changes the cost math for running high-volume agentic classification or extraction pipelines against live product or SERP data.
Our takeThe number worth tracking here isn't the parameter count, it's the benchmark-per-dollar ratio: a 304B model priced under larger, weaker competitors is the same value-compression pattern that has been showing up across frontier labs all year. Whether that holds on a practitioner's own agentic workload, not just a benchmark suite, is the only test that actually matters before a production pipeline switches to it.
Confirmed
MCP's July 2026 spec lets clients and servers skip the initialize handshake entirely by having every request self-declare its protocol version
The Model Context Protocol's current spec revision, dated July 28, 2026, has every request carry its own protocol-version field, so two modern implementations skip a separate negotiation step; a new server/discover RPC lets a client optionally probe a server's supported versions and capabilities in one call. Legacy 2025-11-25-era clients and servers keep working through a documented backward-compatibility path. Developer Simon Willison built mcp-explorer, a CLI defaulting to the new protocol, and a version-aware Datasette endpoint.
In plain termsMCP is the emerging standard that lets an AI assistant plug into outside tools and data sources; this update lets one request establish the connection and start working, instead of a two-step handshake the server has to remember afterward, which makes it easier to run at scale.
Why it mattersAny tool or agent wired to an MCP server needs to know whether it is talking to a handshake-era, pre-2025-11-25 server or a modern one, since the connection code differs and both will exist side by side during the transition.
Our takeStateless protocols scale more predictably because a server does not have to remember who is mid-conversation, which is the same tradeoff web APIs made when REST displaced stateful RPC. MCP integrations built against the handshake era should treat the backward-compatibility window as a deadline rather than a permanent fallback, since deprecated features in this spec carry a stated removal path instead of indefinite support.
Confirmed
datasette-agent adds a browser_task mechanism letting agent tools execute code directly in the user's browser
Simon Willison's datasette-agent project shipped version 0.4a0 on July 31, 2026, adding an await context.browser_task() mechanism that lets an agent tool run code inside the user's own browser session rather than on the server. The release notes describe it as letting Datasette Agent plugins provide tools that execute custom JavaScript in the user's browser, extending what a server-side agent can do without a separate headless-browser dependency.
In plain termsInstead of the AI running its actions on a distant server, this hands a small piece of code to your own browser to run using whatever you are already logged into, similar to a browser extension acting on the AI's behalf.
Why it mattersAn agent that can execute JavaScript in a real, already-authenticated browser session interacts with a page the same way a logged-in human would, a materially different capability and risk model than an agent reading static HTML.
Our takeExecuting arbitrary code in an already-authenticated browser session is a capability jump past merely reading a page's structure, and it is the kind of tool permission that warrants the same two-stage-commit scrutiny as any other irreversible action an agent can take. The open question for a site is less about crawlability now and more about what an agent can do once it is logged in as a real user.