Read Online | Sign Up | Advertise | Ad-free

Presented by

Good morning, AI enthusiasts. The Model Context Protocol and two other agent frameworks just left private control. Under a new Linux Foundation project, MCP, AGENTS.md, and goose are now community-governed open standards. AWS, Bloomberg, Cloudflare, Google, and Microsoft backed it on day one.

In today's recap:

  • Anthropic, OpenAI, Block launch agent standards foundation

  • Study shows 57% of agent use is cognitive work

  • Expert personas don't improve AI accuracy, study finds

  • 4 new AI tools, prompts, and events

AI INFRASTRUCTURE

Tech giants launch Agentic AI Foundation

Linux Foundation

Recaply: Anthropic, OpenAI, and Block just established the Agentic AI Foundation under the Linux Foundation, donating three foundational projects (MCP, AGENTS.md, and goose) to create vendor-neutral standards for agentic AI development.

Key details:

  • The foundation operates as a directed fund under the Linux Foundation, providing neutral governance for open standards including Anthropic's Model Context Protocol, OpenAI's AGENTS.md format, and Block's goose framework.

  • AGENTS.md has been adopted by more than 60,000 open-source projects since August 2025, while MCP now powers over 10,000 active servers across developer tools and Fortune 500 deployments.

  • Platinum members include AWS, Bloomberg, Cloudflare, Google, and Microsoft, with dozens of additional companies joining as gold and silver members to support the neutral infrastructure.

  • The foundation launches immediately with all three projects transitioning to community governance, allowing contributors from any organization to shape future development of agentic AI standards.

Why it matters: The industry faces a critical fork in how agentic AI develops, and the AAIF aims to prevent ecosystem fragmentation. Without neutral standards, agent development risks splitting into proprietary silos where competing tools can't work together, limiting portability and progress. History shows that open infrastructure like the Internet, Linux, and the Web accelerates innovation faster than closed ecosystems, but establishing those standards requires coordination now before competing approaches become entrenched and interoperability becomes impossible.

PRESENTED BY HUBSPOT

How can AI power your income?

Ready to transform artificial intelligence from a buzzword into your personal revenue generator

HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.

Inside you'll discover:

  • A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets—each vetted for real-world potential

  • Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background

  • Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve

Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.

Get Your Guide

PERPLEXITY & HARVARD

Study maps real-world AI agent usage

Perplexity

Recaply: Perplexity and Harvard researchers just published the first large-scale field study analyzing hundreds of millions of real-world AI agent interactions, revealing that 57% of agent activity focuses on cognitive work rather than simple task automation.

Key details:

  • The research analyzed anonymized interactions from Comet and Comet Assistant users to answer three questions (who adopts agents, how intensively they use them, and what tasks they delegate), making it the first study of actual agent behavior at scale.

  • Productivity and workflow tasks account for 36% of all queries while learning and research represents 21%, far exceeding expectations that agents would primarily handle rote chores like booking hotels or managing calendars.

  • Personal contexts drive 55% of query volume compared to 30% professional and 16% educational use, with users shifting from low-stakes queries (travel, trivia) toward complex tasks (debugging code, synthesizing research) as they gain experience.

  • Digital technology workers generate 30% of queries by volume, but marketing, sales, management, and entrepreneurship show higher usage intensity once adopted.

Why it matters: After years of speculation about how AI agents would be used, the data settles a key debate. Users aren't delegating mindless busywork to save time, they're scaling their cognitive capacity to do harder work. This mirrors the personal computer's trajectory from recipe management to spreadsheets, suggesting AI agents are following the same pattern from novelty to necessity for knowledge work.

AI RESEARCH

Expert personas don't boost AI accuracy

Ethan Mollick

Recaply: Wharton School researchers just published findings showing that assigning expert personas to AI models doesn't improve factual accuracy on technical questions, testing six models across 498 graduate-level questions in physics, chemistry, biology, engineering, and law.

Key details:

  • The study tested three persona types (in-domain experts matched to question type, domain-mismatched experts, and low-knowledge personas like "toddler") with 25 independent responses per question across GPQA Diamond and MMLU-Pro benchmarks.

  • Expert personas showed no consistent benefit across models with few exceptions, while domain-mismatched personas sometimes degraded performance and low-knowledge personas (especially "toddler") reduced accuracy across multiple models.

  • Gemini 2.0 Flash on MMLU-Pro was the main outlier showing improvements for all five expert personas, while Gemini 2.5 Flash frequently refused to answer when given out-of-domain expertise, asserting it lacked relevant knowledge.

  • The research used 4,950 runs per model for GPQA Diamond (198 questions) and 7,500 runs per model for MMLU-Pro (300 questions), following previous work showing that 25 trials provide sufficient statistical power.

Why it matters: Persona prompting has been a hot topic over the last few months, with major AI companies recommending the technique in official documentation. While Google, Anthropic, and OpenAI have been pushing persona-based approaches, the Wharton team's rigorous testing may bring new skepticism to an area still waiting for a prompting breakthrough that consistently improves model accuracy on difficult questions.

NEWS

What Matters in AI Right Now?

  • ServiceNow released Apriel-1.6-15B-Thinker, a 15B-parameter multimodal reasoning model scoring 57 on Artificial Analysis Index, reducing token usage by 30%.

  • Mistral released Devstral 2 coding model family with 123B and 24B versions, achieving 72.2% on SWE-bench Verified under modified MIT license.

  • Anthropic updated Claude Agent SDK with 1M token support and security tools for enhanced agent development workflows.

  • Nous Research open-sourced Nomos 1, a mathematical reasoning model scoring 87/120 on Putnam 2025, trained with Hillclimb AI collaboration.

  • OpenAI launched ChatGPT Foundations for Teachers course on Coursera, targeting 10M Americans by 2030.

  • Accenture and Anthropic launched multi-year partnership with 30,000 professionals trained on Claude, forming Accenture Anthropic Business Group for enterprise AI.

  • Microsoft invested $17.5B in India for cloud and AI infrastructure through 2029, doubling skilling commitment to 20M people by 2030.

TOOLS

Trending AI Tools

  • ⚡ Vybe: Build secure internal apps with AI and 3000+ integrations *

  • 🔍 Scouts: Always-on AI agent monitoring evolving information

  • 💻 Vibe CLI: Mistral 2 AI coding agent for multi-file edits

  • 🧑‍💻 Claude Code: Most powerful AI coding assistant, with new Slack integration

* Sponsored

PROMPTS

SEO-friendly Blog Outline 

Prompt: Act as a professional copywriter. Draft me a SEO-friendly blog outline about [Insert Blog Topic] , and target audience who are [Insert Target Audience], with an objective to [Insert Objective] and rank high on search engine result page. Include relevant keywords for each section. Tone is [Insert Tone]. Format in a bullet point.

Source: Grace Leung
EVENTS

🧡 Enjoyed this issue?

🤝 Recommend our newsletter or leave a feedback.

How'd you like today's newsletter?

Your feedback helps me create better emails for you!

Login or Subscribe to participate

Cheers, Jason

Connect on LinkedIn, & Twitter.

Reply

or to participate

© 2025 Recaply, Inc.

Privacy policy

Terms of use

Powered by beehiiv