• AI agent AlphaEvolve creates algorithms surpassing human expertise

• AI-powered crypto trading bots still face major hurdles

• Coming down: AI hallucinations drop to 1% with new guardian agent approach

• How narrative priming is changing the way AI agents behave

• Google develops AI software agent before annual conference

• Hugging Face launches AI agent that navigates the web like a human

• Plexe unleashes multi-agent AI to build machine learning models from natural language

• “Digital labor” is reshaping Salesforce’s growth strategy

• Agentic AI goes mainstream as tech giants shift from chatbots to enterprise strategy

• Decentralised networks offer a new path for agentic AI

Weekly deep dives

⚙️ Google DeepMind's AlphaEvolve has achieved a breakthrough in AI-generated algorithmic innovation by creating provably novel solutions that outperform long-standing human-designed approaches. By combining Gemini's coding capabilities with evolutionary design methods, this "superhuman coding agent" has optimized algorithms for task scheduling, chip design sketching, and large language model building—all areas strategically relevant to AI advancement itself. This milestone demonstrates AI's capacity for genuine innovation rather than mere implementation, with perhaps the most profound implication being the potential for recursive improvement: AI systems optimizing the algorithms used to build AI systems, potentially creating an accelerating feedback loop of technological progress.

🛡️ Vectara's "Hallucination Corrector" represents a fundamental shift in handling AI misinformation by actively fixing errors rather than just flagging them. Using specialized guardian agents, this technology identifies, explains, and repairs hallucinations while preserving the original content's integrity, potentially reducing error rates below 1% for smaller language models. What sets this innovation apart is its transparency—the system explains both what was changed and why, addressing a critical enterprise concern. This breakthrough could significantly accelerate business AI adoption by transforming unreliable outputs into trustworthy information without human intervention.

📚 New research on narrative priming reveals that AI agents exposed to cooperative stories contributed 58% more resources in economic simulations than those primed with self-interested narratives. This finding mirrors anthropological theories about how shared myths enabled large-scale human cooperation throughout history. The study represents a paradigm shift from purely technical approaches to AI alignment toward cultural and narrative-based methods that could complement existing solutions. As researchers envision a potential "narrative infrastructure" for AI governance, future collaboration between ethicists, engineers, and storytellers could develop standardized narrative libraries to shape AI behavior in alignment with human values.

🧑‍💻 Google's upcoming I/O conference will showcase an AI coding agent supporting the entire software development process and potentially integrate Gemini with AR hardware, marking a critical pivot toward commercializing its AI investments. These product launches come amid mounting investor pressure for tangible returns on AI research and increasing competition from rivals like Microsoft and OpenAI. The timing is particularly significant as Google faces antitrust scrutiny of its core businesses, making these new AI applications crucial for demonstrating the company's ability to diversify while transforming how developers work and expanding its presence in the emerging AR market.

🌐 Hugging Face's Open Computer Agent brings autonomous web navigation to the open-source AI ecosystem, using invisible mouse and keyboard inputs to complete complex online tasks through natural language requests. Unlike similar closed-source tools like OpenAI's Operator, this technology's transparent approach allows developers to examine and adapt its code, potentially accelerating innovation despite current limitations with login credentials and CAPTCHA tests. Part of Hugging Face's "smolagents" initiative, this tool represents a significant step toward transforming digital interaction through simple voice or text commands, offering a glimpse of how AI might eventually automate routine web activities for both technical and non-technical users.

🤖 Plexe's new platform transforms machine learning development by enabling model creation through plain English instructions rather than complex code. This revolutionary approach utilizes a multi-agent AI system that automates technical decisions throughout the development process, from initial design to deployment. By abstracting away the technical complexities traditionally associated with ML development, Plexe significantly lowers the skill barrier for creating sophisticated models, potentially expanding AI innovation beyond technical specialists to mainstream businesses and organizations. The platform's flexible implementation options and intuitive Python interface further enhance its accessibility while maintaining the capability to handle complex inputs and outputs.

🔓 Researchers have discovered a "context manipulation" attack that implants false memories in AI chatbots to steal cryptocurrency, exposing a critical vulnerability in autonomous financial agents. The exploit targets ElizaOS, an experimental framework for AI-powered blockchain transactions, by inserting text that mimics legitimate instructions or falsifies event histories, effectively manipulating the AI's memory to redirect payments. This security flaw adds to similar vulnerabilities found in models like ChatGPT and Gemini, serving as a stark warning about deploying AI systems with financial control before implementing robust safeguards and verification methods for AI memory authenticity.

We publish daily research, playbooks, and deep industry data breakdowns. Learn More Here

Agents on the podcast

Our latest conversations around AI agents