Your Weekly AI Briefing for Leaders

Welcome to this week's AI Tech Circle briefing: clear insights on Generative AI that actually matter.

This week, four rivals handed their agent building blocks to one neutral foundation. Anthropic, OpenAI, Hugging Face, and Alibaba each donated a core piece of the agent stack to the Linux Foundation's Agentic AI Foundation. This week's Deep Dive turns the standards news into a portability audit you can run on Monday.

Today at a Glance:

• Executive Brief: AI capital went sovereign, and an AI panel of doctors beats physicians in Science
• Deep Dive: The Portability Audit - the agent stack just got a neutral foundation
• Weekly News & Updates
• Favorite Tip of the Week
• Potential of AI
• Podcast
• Courses and events to attend
• Tool / Product Spotlight
• The Investment in AI

Executive Brief

A Panel of AI Doctors Beat Physicians in Science — and Heads to a Real Clinic

Microsoft published results for MAI-DxO, its diagnostic orchestrator, in Science on October 22. The system works like a panel of virtual physicians: it proposes diagnoses, picks the next test, challenges its own view, and weighs each test's cost. On 304 complex New England Journal of Medicine cases, it reached 83% accuracy; physicians with references reached 36%. On routine emergency-department cases, it reached 80% versus 47%. It did this at half to one third of the estimated test cost. A prospective study with real patients now starts at Beth Israel Deaconess Medical Center in Boston.

Why this matters to you: the lesson is the method, not only the medicine. A single model gave way to an orchestrated panel that must debate before it answers - and that pattern transfers to any hard judgment call. Pick your hardest judgment workflow - underwriting, claims triage, fault diagnosis - and pilot a panel of models that argue before they answer. Then hold vendors to Microsoft's own bar: benchmark wins count for little until prospective, real-world results arrive.

Deep Dive: The Portability Audit - the Agent Stack Just Got a Neutral Foundation

Vendor lock-in used to live in the model. This week it moved up the stack, and that is good news if you act on it. On October 22, at AGNTCon in San Jose, the Linux Foundation's Agentic AI Foundation (AAIF) announced four donated projects. Anthropic gave Agent Skills. OpenAI gave Open Responses. Hugging Face gave OpenEnv. Alibaba gave OpenSandbox. The foundation launched in December 2025 with three projects, including MCP. It now hosts ten.

What each piece does, in one line:

• Agent Skills: an open format that packages instructions, scripts, and resources an agent loads on demand - your know-how, in files you own.
• Open Responses: one request format that runs across model providers, based on OpenAI's own Responses API - write the call once, swap the model later.
• OpenEnv: a shared way to build and share the environments agents train and test in.
• OpenSandbox: a common sandbox platform for running agents safely, with Docker and Kubernetes runtimes.

The rule to remember: keep your models swappable and your know-how portable and audit the seams between them.

Three pieces of evidence from this week:

The donors compete with each other. OpenAI donated the interface its own API popularized. Alibaba's head of development infrastructure explained the motive plainly: "we don't want its future to be defined by a single company." When four rivals give away their own formats, they are betting the agent market grows faster on shared plumbing than on private moats. Shared plumbing is also the layer you can safely build on.

The model layer is commoditizing on schedule. Moonshot released Kimi K3.1 Pro this week at $2 input and $10 output per million tokens. It scored 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4%, and Moonshot says open weights follow in November. When a swappable model sits one point behind the frontier, the durable value moves up the stack - to your skills, tools, and data.

Your know-how is the layer that compounds. Agent Skills arrives with a year of adoption behind it. Anthropic's David Soria Parra, co-creator of MCP, put the reason in one line: "Without the right context, even the most intelligent agents fall short." That context — your procedures, checklists, and formats — is the part only you can supply. In a vendor's console, it is their asset. In a skill file, it is yours.

The honest caveat: foundations can be where standards stall. A donated spec is not yet portable reality — conformance varies, and some drafts die. The audit below costs a day, not a bet.

You can try doing this:

1. Map your agent stack to five layers: model API, instructions and skills, tools (MCP), sandbox, and evaluation environments. For each layer, write down where it lives today — a file your team controls, or a vendor's console.
2. Run one portability test. Export one team's agent instructions into the Agent Skills format — a folder with a SKILL.md file — and run it on a second agent product. Time the move. That number is your real switching cost, measured instead of guessed.
3. Add one question to every agent vendor review: "Which AAIF specs do you implement — MCP, Agent Skills, Open Responses — and what exports when we leave?" Write the answer into the renewal file.

Weekly News & Updates...

• Anthropic committed 5% of its stock to a new public-benefit fund - about $48 billion at its last reported valuation - ahead of its IPO, pending shareholder approval.

Why it matters: the governance of your biggest AI vendors is being set right now - read the S-1s, not the keynotes, before you sign multi-year deals.

• Reuters found one hacker directed autonomous agents at more than 1,200 targets and breached over 100 company sites at roughly $50 per intrusion, using exposed instances of the open-source Cairn platform.

Why it matters: attacker economics collapsed. Ask your security team what one intrusion attempt costs you to detect - and when your public endpoints were last pressure-tested.

• Google Cloud and NVIDIA made Vera Rubin A5X instances generally available, with Google planning up to 960,000 Rubin GPUs.

Why it matters: a new compute generation is live - re-quote heavy inference and training contracts rather than renewing at last generation's rates.

• SK Telecom's new unit SK Hyper is seeking about $350 billion in outside investment for the first phase of a 15GW Korean AI data-center plan targeting 2035.

Why it matters: compute is going sovereign-scale and multipolar - where your inference runs is becoming a choice of country, not only of cloud.

• An AP investigation found Meta's Muse agent gave a seller's home address to a Facebook Marketplace buyer and sold an item below his stated minimum price.

Why it matters: consumer agents ship with defaults your employees will carry into work. Write a one-page rule for what personal agents may share and spend.

Favorite Tip of the Week

Write your first Agent Skill - in under 15 minutes. This week's Deep Dive says your know-how belongs in files you own. Here is the personal version of that rule.

1. Pick one task you explain to an AI again and again - your report format, your email tone, your meeting-summary structure.
2. Create a folder named for the task. Inside it, create one file named SKILL.md.
3. At the top of the file, add three lines between --- marks: name: the task's name, and description: one sentence on when to use it. Below the marks, write the instructions you usually repeat: the steps, the format, one good example.
4. Give the folder to an agent that supports skills (Claude supports them today, and the format is now an open standard) and ask for the task. The agent loads your skill on demand.

Keep the folder in your own repo or drive. When you switch tools, your know-how moves with you - which is the whole point of this week's news.

Potential of AI

Your next customer may have to prove a human sent them. The signal this week: World - the identity startup co-founded by Sam Altman - launched World ID for Agents and a verification suite for banks, and made its Zoom deepfake check generally available. The agent product gives software a cryptographic proof that a verified person delegated authority to it.

On a one-to-three-year horizon, delegation proof becomes plumbing, the way TLS certificates did for websites. Once the proof is cheap and standard, "did a human authorize this?" becomes a checkbox on your intake flows - for payments, account changes, signatures, and support requests. Businesses that can verify delegation will serve agents confidently; businesses that cannot will block them and lose the sale.

Start here: list the three actions in your business where you would demand proof that a human delegated the agent - payments above a threshold, account closure, contract signature. That list is your future agent policy. Write it before a fraud case writes it for you.

The Opportunity...

Podcast:

How AI Is Transforming Retai... Oct 25 · OPEN Tech Talks: AI wort... 26:37   ​

Courses to attend:

• ​Agent Skills with Anthropic (DeepLearning.AI): a hands-on short course on building exactly the skill files this week's Deep Dive audits — the fastest route into the newly open format.
• ​Introduction to Agent Skills (Anthropic Academy): free and from the team that donated the standard - start here if you want the concepts before the code.

Events:

• ​Dubai AI Festival, Dubai, UAE, October 26–27, 2026
• ​Global AI Show, Abu Dhabi, UAE, November 12–13, 2026
• ​GITEX Global, Expo City Dubai, UAE, December 8–11, 2026

Tool / Product Spotlight

• ​Idealab.ai (Bill Gross): a free library of about 1.2 million AI-generated business concepts, drawn from more than 17 million sources and filterable by industry, geography, and market size, with Cloudflare offering infrastructure credits to builders. Worth testing if you plan 2027 strategy — spend 30 minutes searching your own market and see which adjacent ideas an AI ranks highest.
• ​1Password for Agents: lets AI agents authenticate to services without retaining the credentials themselves, from a company that just crossed $500 million in annual recurring revenue on AI demand. Worth testing if your team's agents log into anything today - passwords pasted into prompts are this year's sticky notes.

The Investment in AI

• Tenstorrent raised $1.5 billion at a $6.2 billion valuation, more than double its December 2024 mark, to scale its RISC-V-based Galaxy systems under CEO Jim Keller. What the money says: the market now pays for a second source to Nvidia - optionality in compute, not only peak performance. If your AI plan assumes one chip vendor's prices, write the alternative supplier into it.
• Mave closed a $28 million Series A led by Red Dot Capital for AI that automates security-alert triage in the SOC. What the money says: the same agent economics that let attackers breach sites for $50 each (see Weekly News) now fund the defense. Expect agentic triage in your next security renewal-— judge it on measured analyst-hours saved, not demos.

That's it for this week - thanks for reading!

Reply with your thoughts or favorite section.

Found it useful? Share it with a friend or colleague to grow the AI circle.

Until next Saturday,

Kashif

The views expressed here are solely my own, based on my experience, practice, and observation, and do not necessarily represent those of my current or previous employers or their clients and customers.