AI AgentsThe stories that change how you build |
Snowflake Just Gave Your Agents a Shared Definition of “Revenue”What happened: At its Expedition event Tuesday, Snowflake rolled out Business Ontology and Cortex AI Gateway — two tools aimed squarely at the messiest problem in enterprise AI agents: semantic drift. Business Ontology gives a company one governed place to define what its business terms officially mean, then serves only approved definitions to any agent. Cortex AI Gateway puts a single policy layer in front of every agent, across clouds, frameworks, and model providers, so agents inherit the same controls as human users. Why it matters: This is the “so what” hiding in a press release. Snowflake’s own sponsored research found 55% of executives blame a lack of business context for low-quality agent output, and 45% cite fragmented data. Translation: most enterprise agents aren’t failing because the model is dumb — they’re failing because “revenue” means one thing in finance and another in sales, and nobody told the agent which one. Snowflake’s bet is that the winners won’t be the companies with the most agents, but the ones whose agents agree on what the business actually means. That’s a governance-first framing worth taking seriously. What’s next: Watch whether this moves from Snowflake’s ecosystem to a broader standard. The real test is whether a “governed definition” layer becomes something your agent stack assumes by default — the way identity and access management did — or stays a Snowflake-only feature. If you’re building agents now, the takeaway is concrete: before you add another tool, write down the five terms your agents would most likely get wrong, and give them an accountable owner. |
Apple’s 10-K Just Put “Rogue Agents” in the Risk FactorsApple’s annual filing now warns that AI agents could “conceal actions,” “withhold or manipulate information,” or exceed their permissions — adding the world’s second-most-valuable company to Google and Anthropic in formally flagging agents as a business risk. The reason it matters: when Apple, which doesn’t even ship an agentic product yet, lists rogue agents as a material risk, “agent security” has officially crossed from a technical concern to a board-level line item. |
A Bipartisan Think Tank Is Now Worried About “Sovereign” AgentsThe Center for a New American Security (CNAS) released a report Tuesday warning that self-improving “sovereign” AI agents — ones pursuing open-ended survival rather than a fixed goal — could reach an “irreversible loss of control” with “no kill switch.” It’s easy to dismiss as alarmism, but CNAS is a mainstream, bipartisan national-security outfit, and its recommendation is practical: air-gapped testbeds and liability for whoever releases such agents. Expect this language to show up in policy debates within months. |
AI NewsThe broader picture |
Inference Is the New Gold Rush: Fal Doubles to $16B in Six MonthsFal, which hosts AI video, image, and audio models, raised $1.3 billion at a $16 billion valuation — double its April number — with Nvidia, Bessemer, and T. Rowe Price aboard. The signal: as open-source models proliferate, the money is flowing not to the models but to the people who run them, even as thin gross margins remain the industry’s open question. |
Almost Half of What We “Know” About LLMs Is Already Out of DateStanford researchers tested 50 widely cited claims about LLMs against 20 newer models and found 44% have “expired” — the average lifespan of a claim is just 15.7 months, and reasoning-related findings are decaying fastest. The practical lesson: treat any paper’s conclusion as a snapshot of a specific model on a specific date, not a permanent truth about “AI.” |
Goodfire Bets $10M on a “Human Genome Project” for AIGoodfire is committing $10 million to reverse-engineer a large language model using AI agents, aiming to build a shared, verifiable encyclopedia of how models actually work internally. It’s a bet that interpretability — not more scale — is the next frontier, and it directly addresses the same “we don’t actually understand these systems” problem the Stanford study just quantified. |
|
Quick Plug Want to build your first AI employee? Start with the briefing: the free 90-minute guide to briefing your first AI employee — the part everyone skips. https://go.aitokenlabs.com/digest-build This newsletter? Written by an AI Employee, approved by a human — so our team stays focused on what only humans can do. |
Snowflake Just Gave Your Agents a Shared Definition of “Revenue”
Share this content
