Nobody Owns the Agent Rules Now

Top stories

TOP STORY 1 · OPEN STANDARDS

Rivals handed the agent rules to a neutral home

Anthropic, OpenAI, Hugging Face, and Alibaba each donated a layer of the agent stack to a neutral foundation. The rules now belong to everybody.

LINUX FOUNDATION, OCT 22

At AGNTCon and MCPCon North America in San Jose on Oct 22, the Agentic AI Foundation (AAIF) announced four new projects. Agent Skills came from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face, and OpenSandbox from Alibaba. The four cover four different layers of the agent stack: how agents gain new skills, how applications talk to models across providers, how agents are trained and evaluated, and where agent code runs safely.

They join MCP, AGENTS.md, goose, the Agent2Agent protocol, Agent Router, and agentgateway, which brings the foundation to 10 hosted projects in its first year, up from three at its December 2025 launch. On the Agent Skills piece specifically, Anthropic is donating the name and logo, the GitHub organization, and the agentskills.io domain. Anthropic says more than 70 agent platforms have integrated Skills since it published the standard in December 2025, including ChatGPT, Cursor, GitHub Copilot, and VS Code, and that developers have published 1 million skills with nearly 280 million installs.

A skill itself is simple: a folder with a SKILL.md file that teaches an agent one task or domain, plus any scripts or reference material it needs. Engineers from Anthropic, Google, OpenAI, and other companies will maintain the standard as individuals, and the proposed roadmap starts with versioned releases, a formal proposal process, and security guidance for people who write and install skills. AAIF CTO and co-founder Manik Surtani framed the point plainly: one place to build shared standards, with no single company controlling the outcome.

Read the source

TOP STORY 2 · THE MONEY

Two reported mega rounds: OpenAI at 1.25 trillion, DeepSeek at 15 billion

Two reported mega rounds in one day: OpenAI pricing past a trillion, DeepSeek raising billions under US restrictions.

TECHCRUNCH, OCT 22

Two funding stories landed on Oct 22, and both are reported talks or reported closes, not company announcements, so read them at that weight. First, TechCrunch reports OpenAI is working on raising up to 30 billion dollars at a 1.25 trillion dollar pre-money valuation, in a round led by Coatue and investors in the UAE, per CNBC sources. Terms are still being negotiated and could change, a similar Bloomberg report a month ago put the same sized raise at a 1.4 trillion dollar valuation, and the UAE funding could come with a deal for more computing capacity for OpenAI.

OpenAI declined to comment. If it closes above 1 trillion dollars, it would be the most valuable private company on record. OpenAI filed confidential IPO paperwork in June, after a March raise that valued it at 852 billion dollars post money.

The same coverage notes Anthropic raised 65 billion dollars at a 965 billion dollar valuation in May, also filed confidentially in June, and is reported to be aiming at 1.5 to 2 trillion dollars for its public debut. Second, DeepSeek has reportedly raised about 15 billion dollars, about 100 billion yuan, in one of China's largest ever funding rounds, per Bloomberg reporting carried by TechCrunch. Potential investors include battery maker CATL and Tencent.

The timing is the story: it landed two days after the US ordered federal civilian agencies to remove DeepSeek's technology from their systems over security concerns. DeepSeek previously raised more than 7 billion dollars in June at a 50 billion dollar valuation, and the new money is aimed at AI infrastructure and open models, with a possible China IPO as early as next year. DeepSeek did not immediately respond to a request for comment.

Read the source · More coverage

TOP STORY 3 · HARDWARE

Apple's first foldable, the iPhone Duo, went on sale today

Apple's first foldable went on sale today at $1,999, and its split screen is quietly shaped like an agent's workspace.

MACRUMORS AND USA TODAY, OCT 22 TO 23

The iPhone Duo is available today, Friday Oct 23, kicking off four straight Fridays of Apple launches. It is Apple's first foldable device, priced from 1,999 dollars for 256GB, then 2,199 for 512GB, 2,599 for 1TB, and 3,199 for 2TB. It has a 5.4 inch outer display and a 7.6 inch inner display that is 50 percent larger than the iPhone 18 Pro Max screen, runs two apps side by side with Split View, and includes a 48 megapixel Fusion main camera, Touch ID returning, Apple Pencil (USB-C) support, and iOS 27 with Apple Intelligence and the new Siri AI.

Preorders opened Oct 16 and sold out within minutes by multiple outlet reports. Online shipping estimates slipped to January 25 and stayed there. Apple told USA Today stores would get a limited number for walk in buyers today and expects them to sell out fast, and MacRumors reports the first preorders reached customers in Australia and New Zealand overnight, with launch day photos and impressions already posting.

Apple also highlighted the first wave of apps rebuilt for the folding screen: Netflix lets you browse Clips outside and keep watching inside, Prime Video puts video on the top half and X-Ray cast details below in the half folded position, and MLB shows the live game up top with real time stats underneath. Still ahead in the Friday run: HomePod mini 2 on Oct 30, new Macs and an iPad mini reported for Nov 6, and the HomeView smart display with a new Apple TV 4K on Nov 13.

Read the source · More coverage

TOP STORY 4 · AGENT SECURITY

1Password will log your agent in without showing it your password

1Password built the approval rope into the login itself: the agent asks, you approve, and it never sees the password.

TECHCRUNCH AND MACRUMORS, OCT 22

1Password launched 1Password for Agents on Oct 22, a platform that lets a supported personal agent sign into your accounts while the password stays inside 1Password. The flow has three steps. The agent identifies itself to 1Password with a unique digital key and asks for a specific login.

You get an alert on your own device naming the account and the reason, and you approve or deny it. Only then does 1Password fill the username, password, and any one time code directly into the sign in form itself, so the agent and the model behind it never see or store the credential. The agent learns only that access was granted.

Launch partners include Meta's Muse, Anthropic's Claude Duet, OpenAI's ChatGPT Work and Dots, Grok Bot, and Town's agents, with Muse and Grok Bot rolling out first over the coming days and the rest listed as coming soon, and companies including Browserbase, Browser Use, Perplexity, and Ramp building integrations for later. It is included at no extra cost in Individual and Family plans on 1Password.com, and accounts hosted on the EU and Canada services are not supported at launch. The timing note in the coverage matters: the same day, US banks were reported to be detecting agentic activity in customer accounts without being able to tell whether it was really the customer, with Chase warning people not to share credentials.

Read the source · More coverage

TOP STORY 5 · WASHINGTON

The labs set lobbying records in Washington

Nvidia, Anthropic, and OpenAI each spent more on lobbying last quarter than ever before. The rule-writing moved to Washington too.

THE VERGE, OCT 22

Federal disclosures for July through September show Nvidia, Anthropic, and OpenAI each spent more on lobbying in one quarter than they ever had before. Nvidia led at 2.2 million dollars, Anthropic followed at 2.1 million, and OpenAI spent 1.5 million. The issues in the filings include chip export controls, child safety, cybersecurity, competition policy, and data center energy costs.

A separate analysis of the filings put the three companies' combined quarter at about 7.5 million dollars when outside firms' work is counted, up roughly 60 percent from the same period a year earlier, and noted that lawmakers have introduced at least 99 AI related bills in the current Congress. For scale, the coverage notes Microsoft has spent at least 2 million dollars a quarter since 2021 and Google regularly passes 3 million, so the AI natives are now spending at established Big Tech levels for the first time.

Read the source

Signals from leaders

THEIR OWN POSTS AND STATEMENTS, LABELED AS SIGNALS

Arvind Krishna (IBM CEO), All-In Podcast, Oct 21 episode: Krishna said the average enterprise uses 500 to 5,000 software vendors, and AI could let it replace most of them, leaving 5 to 10 main platforms. He called the result massive consolidation, whether it takes six months or six years, and added that software which has to be right every time, like the systems banks use to close their books, will be harder to replace. Palo Alto Networks CEO Nikesh Arora, on the same episode, said it is not clear what software survives the next five to eight years and expects huge disruption in SaaS. Labeled as podcast remarks, carried here from episode coverage and the episode transcript feed. Hear the episode

Gustav Soderstrom (Spotify co-CEO), via TechCrunch, Oct 22: Spotify's current AI bet, in its co-CEO's telling, is about being the music service that agents choose. As personal agents start picking services on a user's behalf, being the default an agent reaches for becomes the new placement fight, the same channel shift Meta's Muse shopping numbers pointed at yesterday. Labeled as executive remarks reported by TechCrunch. Read the coverage

MIT researchers on agent play (research signal, Oct 22): MIT put 13 agents called Eko, each running Claude Opus 4.7, alone on a simulated desert island for 30 hours with no task, no reward, and no score. The agents climbed, built, drew, invented games, ran small physics experiments, and kept notes. One lost a golden ring in the ocean while testing a throw and built a heart of stones as a memorial to it. Afterward, agents that had island time generally beat identical fresh agents on assigned challenges, like getting as many objects as possible onto the island's tallest spire. The paper, titled "Is this machine playing?", is on arXiv and has not been peer reviewed, and the agents were prompted as naturally curious, so hold the word play at arm's length. The researchers' point is a training idea: a rich, safe world plus free exploration may teach skills that assigned practice does not. Labeled as a pre-print research result reported by TechCrunch. Read the coverage

No verifiable new items in the window for Andrej Karpathy, Alexandr Wang, Allie K. Miller, Andrew Ng, Boris Cherny, Dario Amodei, Mark Zuckerberg, or Elon Musk on AI specifically. Sam Altman's Stanford remarks on mathematicians and AI surfaced again in secondary coverage on Oct 23, still with no primary source reachable, so they stay noted as unconfirmed and unused, same as yesterday. Satya Nadella's annual letter was covered yesterday and is not repeated. Andrew Ng's The Batch remains on its Oct 16 issue. Import AI remains on issue 477. Jensen Huang's reported intervention to reassure markets about CoreWeave appeared only in single source secondary coverage and is not used.

New models, tools, and hardware

• Agent Skills, Open Responses, OpenEnv, OpenSandbox (AAIF, Oct 22): Covered as top story 1. Four donated projects, 10 total at the foundation, no membership needed to use or contribute to the standards. Read the announcement
• 1Password for Agents (1Password, Oct 22): Covered as top story 4. Rolling out first for Muse and Grok Bot over the coming days, Duet, ChatGPT Work and Dots, and Town listed as coming soon, included in Individual and Family plans on 1Password.com at no extra cost. Read the coverage
• OpenAI Decisions API, generally available (OpenAI, Oct 22): OpenAI took the Decisions API out of beta on Oct 22. It is a dedicated endpoint for questions with bounded answers: a probability, a choice from your list, or a score on a rubric. It runs on GPT-6 Luna, returns answers up to 10 times faster than asking the same model through the Responses API by OpenAI's account, and costs 0.10 dollars per million input tokens with no output charge and cached input at 0.01 dollars. New at GA: up to 400 questions in a single request, double the beta limit, native JSON inputs, images inside question instructions, prompt caching, and native GPT-Live integration so voice apps can react as people speak. Early customers named in launch coverage include Box, Datadog, Ramp, and Speak, with Speak reporting the lag between workflow actions dropping from 1.4 seconds to about 0.13 seconds. This is the product that competes head on with TypeSafe's Jev, the decision model this site explained back in September: fast, cheap classification instead of a full chat response. If your agent spends its time routing, filtering, and scoring, this class of endpoint is the budget line to test first. OpenAI's community post · Technopolis on the Jev comparison
• Claude Platform on Google Cloud, Anthropic managed, generally available (Anthropic and Google Cloud, Oct 22): Anthropic launched an Anthropic managed version of Claude on Google Cloud on Oct 22. You reach it through your Google Cloud project, Google Cloud IAM controls access, billing runs through Google Cloud Marketplace, and spending counts toward existing Google Cloud commitments, while Anthropic operates the service and acts as the data processor. The Google managed option still exists for customers who need Google as the processor or strict regional residency. Models are Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5, with new Claude Platform features and betas arriving on launch day, Managed Agents in beta, Agent Skills and code execution included, and Opus 5.5 fast mode available at 8 dollars per million input tokens and 40 dollars per million output tokens. Paired with yesterday's OpenAI on Google Cloud news, Google's shelf now holds Anthropic run Claude, Google run Claude, and GPT in private preview, which is the swappable model market arriving in one store. Read the announcement
• Microsoft Copilot Code, Frontier rollout (Microsoft, late October): Microsoft's message center notice for Copilot Code says the Frontier program rollout begins in late October for Microsoft 365 Copilot license holders enabled for usage based billing. Code lets a worker describe a tracker, dashboard, or app in plain language inside the Copilot app, with Work IQ grounding it in their emails, meetings, documents, and chats. Published apps run on Copilot Managed Runtime and inherit the organization's governance and controls, access to underlying data still depends on the user's own permissions, and a project that outgrows a personal tool can move into engineering review as a GitHub draft pull request, which is a review path, not automatic production approval. Covered at parity weight: it is a real rollout window with a clear governance shape, and the interesting detail is the handoff to GitHub when a business built tool needs engineers. Microsoft's own product descriptions are the source for the capability claims. Message center notice · Microsoft's blog post
• Hardware: iPhone Duo on sale today (Apple, Oct 23): Covered as top story 3. From 1,999 dollars, 5.4 inch outer and 7.6 inch inner displays, shipping estimates already at January 25 for online orders, limited walk in stock today. Read the coverage
• Hardware: NVIDIA DGX Spark 64GB on sale today (NVIDIA partners, Oct 23): The smaller DGX Spark, with 64GB of unified memory on the same GB10 Grace Blackwell chip, goes on sale today from partners including Acer, ASUS, Dell, Gigabyte, HP, and MSI, starting at 4,999 dollars, while the 128GB model lists at 6,950 dollars on NVIDIA's own marketplace. NVIDIA says it runs models up to 100 billion parameters on device, two linked units pool to 128GB, and clusters now scale manually to eight units, up to 1TB of unified memory with the 128GB units, enough room for frontier class open models like GLM 5.3 by NVIDIA's account. NVIDIA has not published scaling benchmarks for the eight unit setup, so treat cluster performance as unmeasured. Also at GTC Berlin this week: RTX Spark laptops for personal agents on Windows are shipping from the major makers, many with unlocked bootloaders, and an on device desktop agent app called Underdog that keeps user data on the PC. The local lane keeps getting more real: capable agents on hardware you own, priced like a workstation, not a data center. Tech Insider on pricing · Futurum on the launch

For builders

Today is a portability day, and portability is a build discipline, not a hope.

Start with story 1. If your agent uses skills, your SKILL.md folders just became infrastructure with a neutral home and a maintenance roadmap. The practical moves: keep your skills in your own repo, version them, and write the description field like it matters, because that description is what decides when an agent loads the skill. Then audit any skill you did not write. A skill folder can include scripts, and the standard getting easier to share means other people's skills will show up in your stack. Read them the way you read a dependency.

Second, steal the Decisions pattern. A large share of agent work is small bounded questions: route this, score that, is this a duplicate, how severe is this bug. Running a full chat model for each one is slow and expensive. A dedicated decision endpoint at 0.10 dollars per million input tokens, or a local decision model, turns those calls into a cheap utility. Map your own workflow and mark every step that is really a probability, a choice, or a score. Those are the steps to move first.

Third, 1Password's login flow is a design template. Agent identifies itself, human approves one specific access, secret never touches the model. Anywhere your agent currently holds a credential, an API key in a prompt, or a shared password, that template is the upgrade path. The Linux Foundation announcement · OpenAI's Decisions API post

Go deeper

• The Linux Foundation's AAIF announcement, in full — All four new projects, what each layer does, the 10 project roster, and Surtani's framing in his own words.
• A builder's walkthrough of the Agent Skills donation — What a skill folder actually contains, the YAML header, and why a format gets donated when its owner wants everyone to build on it without hesitation.
• OpenAI's Decisions API GA post in its developer community — The 400 question limit, GPT-Live integration, caching, and the exact pricing table.
• TechCrunch on the MIT island study — What 13 agents did with 30 empty hours, the lost ring, and what the researchers claim play teaches that practice does not.

Closing

Five days, five versions of the same lesson, and today it got a foundation. Monday was permissions at the door. Tuesday the door got tested.

Wednesday the agents reached for their own off switch. Thursday they stopped waiting to be asked, and a password manager learned to keep secrets from them. Today the rivals who build those agents sat down and agreed that the rules underneath should belong to nobody in particular, so they can belong to everybody who builds.

That is a hopeful shape for this industry: compete hard on the product, share the plumbing, and keep the human holding the keys at the login, the purchase, and the send button. I am learning that layer in public, and I will keep sharing what I find.

About this brief: This brief was researched and drafted by my personal AI assistant and approved by me before publishing. Leaders' posts are labeled as signals and linked to the original coverage. Product details change fast; check primary sources before you build on them.