## Sources

1. [This Week in AI: Rethinking the Agent Harness](https://www.oreilly.com/radar/this-week-in-ai-rethinking-the-agent-harness/)

---

### **This Week in AI: Rethinking the Agent Harness** by Michelle Smith

**Main Arguments**
*   **The Primacy of the Harness:** A central argument of the source is that the "harness"—the framework and skills built around an AI model—is now more important for performance and utility than the specific underlying model chosen [1, 2].
*   **Security as a Policy Crisis:** The discovery of deep security vulnerabilities by AI models has shifted the conversation from technical capability to urgent national policy and regulation [3, 4].
*   **Betting on Infinite Scaling:** Massive infrastructure investments, such as data centers larger than Manhattan, signal that major tech companies are pricing in a future where AI model capability continues to scale indefinitely [2, 5].
*   **Democratization of Agent Control:** The "age of the unharnessed agent" allows domain experts with no coding experience to manage and "program" AI behavior using natural language, shifting power away from traditional software developers [6].

**Key Takeaways**
*   **The Mythos Security Discovery:** Anthropic's internal model, Mythos, successfully identified thousands of unknown security vulnerabilities in major operating systems and financial infrastructure, including a bug that had persisted for 27 years [3].
*   **Project Glasswing:** Rather than releasing Mythos publicly, Anthropic launched Project Glasswing, a monitored program for trusted partners to focus on defensive patching [3].
*   **Regulatory Shifts:** The US government is considering moving away from a "light-touch" AI policy toward a review process modeled after the FDA's handling of drugs [4].
*   **Infrastructure Explosion:** Huge physical footprints are being dedicated to AI, including the Stratos project in Utah, which spans 40,000 acres and is planned for 9 gigawatts of power—equivalent to nine nuclear reactors [5].
*   **Agent Evolution:** AI product development has progressed through several phases: document completion, chat loops, tool-calling agents, DAG-based workflows, and the current "harness era" [2].
*   **Harness vs. Model Performance:** Stanford research confirms that the performance gap between a bare model and one with a well-designed harness often exceeds the performance gap between competing models themselves [7].

**Important Details**
*   **Computing Power Deals:** Anthropic recently leased the entire xAI Colossus 1 supercluster (200,000+ GPUs) and has contracts for 3.5 gigawatts of capacity from Google and Broadcom coming online in 2027 [5].
*   **Open Agent Protocol:** Guest John Berryman proposed a future standard called the "open agent protocol," which would allow AI agents to receive specific skills and carry context as they move between different environments, such as from a personal notebook to Wikipedia [7].
*   **Domain Expert Empowerment:** By building skills-driven agents, experts can read and fix agent behavior described in plain English, allowing them to troubleshoot issues directly rather than complaining to developers [6].
*   **The Stratos Project:** This data center deal in Box Elder County, Utah, was approved despite local protests and will be more than twice the size of Manhattan [5].
*   **Upcoming Events:** O'Reilly’s *This Week in AI* series continues with future episodes focusing on cutting through headlines to identify what drives actual business value versus what merely looks good in a demonstration [8].