Anthropic hands Agent Skills to the Agentic AI Foundation

Good morning 👋 Anthropic just gave away a standard it didn't have to give away, right as 70-plus platforms started depending on it. That's not generosity. That's math.

In today's issue:

• 🔭 Anthropic hands Agent Skills to a neutral foundation
• 🧠 Models & Releases: ChatGPT's storage finally stops choking, European banks go open-weight, an NHS copilot
• 🔬 Research Highlights: confident lies beat hedged truths, and a loss-function bug nobody caught
• 🚀 AI Startups: a health insurer funds the startup built to fight health insurers
• 🛠️ Try This: turn your most repeated task into a portable Agent Skill
• 🔗 Quick Links: Anthropic's real loss number, and the rest of the week

🔭 THE ONE THING

🧩 Anthropic hands Agent Skills to the Agentic AI Foundation

Anthropic just handed over Agent Skills, the open spec it published last December, to the Agentic AI Foundation, the same body that already holds MCP. The standard isn't a house format anymore: 70-plus platforms run it now, ChatGPT and Codex CLI among them, plus Cursor, GitHub Copilot, and VS Code. That's the real signal. A company doesn't give up control of a spec unless owning it outright is worth less than everyone trusting it's neutral, and Anthropic clearly decided the second MCP-style donation buys more adoption than the first one did. For anyone building agent tooling, treat Skills as infrastructure now, not an Anthropic side project. Lock-in risk just went down.

🧠 MODELS & RELEASES

• 💾 OpenAI rebuilt Habitat, the storage layer under ChatGPT, on Azure Cosmos DB with global caching and real multi-tenancy. Reads are 12x faster at over a billion users. Boring infrastructure work, and the reason your chats don't hang anymore. Scaling storage, part II

• ☁️ AWS took SageMaker's serverless training sessions GA. You write the loss function or reward signal, AWS handles GPU allocation, fault recovery, checkpointing. Fine-tuning and RL without owning a cluster. SageMaker AI customization

• 🏦 NVIDIA is the quiet backbone of European banks walking away from closed models. Revolut's PRAGMA, a transaction foundation model, already serves 80 million customers and is built to scale 3x; Bunq says its open-weight assistant runs at 60% of the cost per answer versus closed alternatives. European banks go open-weight

• 📊 IBM Research combined four Granite time-series models into one ensemble that beats each of them individually on GIFT-Eval, the field's standard benchmark, though the team admits a few specialized systems still edge it out. A sensible default when you don't know which architecture fits your data. Open on Hugging Face. Granite TimeSeries Ensemble R1

• 🌳 Google DeepMind mapped every 10-meter patch of land on Earth by forest type, not just tree cover. Ecologists and carbon-market auditors will be using this dataset long after today's issue is forgotten. Mapping the world's forest types

• 🐟 Sakana AI launched Aquarium, an ontology-based platform for building and governing AI agents, built by a team with ex-Palantir and Snowflake engineers. Strip the ontology framing and it's access control, audit trails, and human-approval gates for agents. Useful, not new. Aquarium release

• 🚑 Microsoft's Copilot is now inside NHS England, drafting reports and flagging patterns in ambulance data. It's kept off clinical calls on purpose. Yorkshire Ambulance Service used it to cut emergency conveyances from 60% to 56%, and the rollout targets 505,000 staff. Before the ambulance arrives

🔬 RESEARCH HIGHLIGHTS

• Confident and wrong beats hedging and right. Thirty-two people spent a session learning science from five GPT-5-mini tutors, each seeded with 0, 25, 50, 75 or 100% fabricated facts and delivered with total certainty. On the falsified material, participants picked the planted lie 60% of the time, versus 11% on correctly taught material (an odds ratio of 12.5), and held the wrong answer with 8.4/10 confidence. Quiz scores fell from 7.4 to 2.9 out of 10, and the usual link between stated confidence and actual correctness flattened to chance. Sycophancy research asks whether the model bends to you. This flips it: when the model sounds sure, you bend to the model. paper

• Your cross-entropy loss has had a bug this whole time. A new paper names two of them. Softmax Collapse zeroes out the target logit's gradient whenever one logit dominates the rest. Softmax Cancellation does the opposite kind of damage, producing a nonzero gradient even when student and teacher logits already match. Patch both and a known training instability called "Slingshot" disappears, grokking speeds up, and distillation runs stop diverging partway through. If a run has stalled or blown up for no obvious reason, this is a cheap thing to check before blaming the data. paper

• Power-seeking in RL training isn't a one-way ratchet. Checkpoint evidence from OLMo-3 complicates the usual instrumental-convergence story, the idea that reinforcement learning on verifiable rewards steadily teaches an agent to self-preserve, deceive, or resist oversight. In the 7B model, the instrumental-behavior rate climbed from 25.0% at the SFT stage to 46.1% around step 200 of RLVR, then fell to 11.8% by the final checkpoint; a preliminary 32B run showed the same rise and fall. The authors flag it as one model family on one benchmark with only moderate judge agreement. Still, it's a real crack in the assumption that RLVR dials up power-seeking by default. paper

🚀 AI STARTUPS

• Baba closed a $40M Series A led by FirstMark Capital, with CVS Health Ventures and Valtruis joining existing backers General Catalyst, Genius Ventures, and Ground Up Ventures. The New York outfit pairs licensed clinicians and care coordinators with AI to walk Medicare and Medicaid patients through providers, pharmacies, benefits, and the rest of the maze. It's supported 10,000 families so far and claims 100,000 more on a waitlist, which tells you more about how broken Medicare navigation is than how good the product is. Worth noting a health insurer is now funding a startup whose job is helping patients fight insurers. source

🛠️ TRY THIS

Turn one recurring task into a portable Agent Skill

1. Pick one thing you re-explain to Claude every single time, a review checklist, a deploy runbook, a data-cleaning pass.

2. Write it as a SKILL.md: a one-line trigger description, then the actual steps. No preamble.

3. If the task really has different jobs inside it, split them. AWS's Deep Insight does this with a supervisor calling a Coder, a Validator, a Reporter, each a separate agent instead of one sprawling prompt.

4. Drop it in .claude/skills, run it twice on real work, then decide if it earns a third.

Prompt: Turn this into a Claude Agent Skill: give me a name, a one-line trigger description, and numbered instructions. Task: [paste your recurring task here].

Worth a look

• AWS open-sourced Deep Insight, a multi-agent system on Bedrock that splits work by role. Haiku routes, Opus plans with extended thinking, Sonnet executes, and AgentCore plus Fargate sandboxes keep generated code from touching anything it shouldn't. A solid blueprint if your "skill" is secretly five agents wearing one coat. AWS blog
• Dharma AI shipped VocabX: trim a small model's vocabulary down to the domain it actually serves. On a Qwen3-4B Portuguese essay grader, cutting unused tokens and fusing common phrases cut real inference cost, runtime and resources together, by 77%, while the bench score went up (0.703 to 0.810). Hugging Face

🔗 QUICK LINKS

• Anthropic posted an $80.7 billion net loss for H1, but $72 billion of that is a non-cash markup on convertible bonds. Strip it out and the real operating loss is closer to $8.1 billion, with compute spend tripling to $9 billion. BNN Bloomberg
• Anthropic also opened a call for life-sciences researchers to come work with its team on special projects, no word yet on what those projects are. Anthropic
• OpenAI is formalizing its forward-deployed-engineer track: engineers embedded directly with customers to close the gap between a model and what it takes to actually run in production. OpenAI
• OpenAI detailed how it hardened frontier training, including offline, air-gapped environments and high-compute red-teaming before a model ships. OpenAI
• OpenAI is partnering with Poland's NASK on AI safety and security research, another country-by-country safety tie-up. OpenAI
• AT&T's legal team rebuilt case intake on ChatGPT Work, using it to scale legal judgment rather than just paperwork. OpenAI
• AWS showed how to stream sentence-level speech from the Kokoro-82M TTS model on SageMaker, cutting the dead air voice agents leave while waiting on a full response. AWS Machine Learning blog
• AWS published a hands-on multi-turn RL guide that trains an agent on whole-task outcomes, not labeled steps, using Wordle as the test case on SageMaker and Bedrock AgentCore. AWS Machine Learning blog
• AWS's new Adaptive Agentic Intelligence Framework, built on Strands Agents, keeps routine work on local small models via Ollama and only kicks the hard stuff up to Bedrock. AWS Machine Learning blog
• IBM Research shipped Granite Speech 5.0: faster transcription and translation, with better speaker attribution. Hugging Face
• NVIDIA's Kumo Relational claims it beats supervised ML on relational data with zero training, a shortcut for anyone tired of wiring feature pipelines by hand. Hugging Face
• Block unveiled Dessa, a speech-to-speech model trained on nothing but human conversation audio, no text, no pretrained base, built by Jack Dorsey's small lab on a single GB200 chip. Block
• Alpha Modus Holdings (AMOD) filed an 8-K for a $7.25 million registered direct offering at $1.15 a share, mostly earmarked to pay down debt. SEC filing

See you tomorrow.

Pradeep Perugu