The biggest story of the past year has been the rise of AI agents, even among consumers. The speed of AI’s rise set off alarms. Fears that AI would replace traditional software triggered a “SaaSpocalypse” selloff earlier this year, while others worried it would eat into labor budgets.
But our data tells a different story: Far from being a zero-sum game, AI is creating new value across the economy, expanding markets far more than it’s displacing existing revenue. — Read More
Recent Updates
The Real Fight in Agentic Commerce Is Over Who Holds the Card
Meta says shopping is the top use case for Muse, its personal AI agent. Per Benedict Evans, a quarter of early Muse users make e-commerce purchases averaging $200, and Meta plans to monetize through commissions. TLDR Crypto reported that agentic purchases through Stripe’s Link grew 38x in a single month.
So the demand question is settled. What is not settled showed up in four different newsletters this week, from four different angles: when an agent spends money, whose credential is it using, and who gets to see the receipt?
That question is worth more than any model benchmark in commerce right now. — Read More
LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms. This might be a big deal for both optimization and interpretability.
… Throughout the history of deep learning, we encoded our simplistic inductive biases into tiny symbolic optimizers like Adam or Muon. This is changing: our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights. — Read More
Conway’s Law for AI Agents: Rethinking Microservices in the Age of AI
I’ve been spending a lot of time working with AI coding agents on larger codebases, and I’ve started to rethink something I thought I’d already settled years ago: microservices.
For a long time, my advice to smaller engineering teams has been pretty simple. Start with a monolith. Keep things modular, establish good boundaries, and don’t introduce distributed systems complexity until you have a real reason to. I’ve seen enough companies go all-in on microservices to know how that story can end.
But working with AI agents has made me realize there might be another reason to reach for microservices. It has less to do with scaling infrastructure or engineering teams and more to do with managing the complexity and risk of making changes to large systems. — Read More
Agentic Design Patterns: A System-Theoretic Framework
With the development of foundation model (FM), agentic AI systems are getting more attention, yet their inherent issues like hallucination and poor reasoning, coupled with the frequent ad-hoc nature of system design, lead to unreliable and brittle applications. Existing efforts to characterise agentic design patterns often lack a rigorous systems-theoretic foundation, resulting in high-level or convenience-based taxonomies that are difficult to implement. This paper addresses this gap by introducing a principled methodology for engineering robust AI agents. We propose two primary contributions: first, a novel system-theoretic framework that deconstructs an agentic AI system into five core, interacting functional subsystems: Reasoning & World Model, Perception & Grounding, Action Execution, Learning & Adaptation, and Inter-Agent Communication. Second, derived from this architecture and directly mapped to a comprehensive taxonomy of agentic challenges, we present a collection of 12 agentic design patterns. These patterns – categorised as Foundational, Cognitive & Decisional, Execution & Interaction, and Adaptive & Learning – offer reusable, structural solutions to recurring problems in agent design. The utility of the framework is demonstrated by a case study on the ReAct framework, showing how the proposed patterns can rectify systemic architectural deficiencies. This work provides a foundational language and a structured methodology to standardise agentic design communication among researchers and engineers, leading to more modular, understandable, and reliable autonomous systems. — Read More
Castle Walls review – Prime’s first fully AI-generated show is here, and it’s a Vaseline-smeared mess
Just as the world reels from Elon Musk’s AI-powered The Odyssey remake – though “remake” feels rather generous for a seven-minute splat of synthetic sludge – another bucket of AI gunk is dropped on to our faces. For Prime Video’s AI-generated TV series Castle Walls, produced by Turkish production studio Ay Yapım using a melange of generative video tools, has arrived, billed as the first end-to-end AI-created series to premiere on a major streaming platform.
It’s certainly more epic than Musk’s production, consisting of two 17-minute episodes. Set in the 11th century, the show revolves around the historic Alamut Castle, although it’s entirely absent of any real sense of time or place. I was briefly intrigued by the opening images, with their bold palette of blues and mustard browns and yellows – but as soon as the “people” appeared, looking like Vaseline-smeared Madame Tussauds creations slowly melting in the sun, everything came crashing down. — Read More
Practical Loop Engineering
… You’ve probably heard about loop engineering. I talked about it a couple of months ago when I wrote a big blog post about it.
A loop is an autonomous, self-correcting feedback cycle where an AI agent repeatedly acts, tests its results, and adjusts its approach until a specific goal is met.
There are now basically two core primitives you can think about. In Claude Code you have a goal primitive, which can drive a single bounded task forward until you’ve got a particular goal, like a measurable finish line that’s been met. And then loop reruns on a timer or a fixed interval, so you can use it to schedule changes. — Read More
Let Decision Models Watch Your Agents
LLM agents can fail, and their actions can have real consequences.
Just like other automated systems, they need condition monitoring to detect anomalous behavior accurately and promptly.
Decision models, which are specially trained AI models that take relevant context and make quick judgments, make real-time agent anomaly detection much more feasible. — Read More
Anthropic’s moat is Claude
The bull case for Anthropic’s IPO.
Anthropic recently filed for IPO. At an estimated $1.5 to $2 trillion, it will be one of the biggest IPOs of all time, rivaling SpaceX, and possibly the biggest. For reference, Facebook’s valuation when it had its blockbuster IPO in 2012 was $104 billion, and Google’s valuation in 2004 was $23 billion. Anthropic will instantly become one of the world’s most valuable companies — Google itself is now worth $3.4 trillion, and Nvidia and Apple are both worth around $5 trillion.
This naturally has a lot of people asking: What’s the upside to buying Anthropic? — Read More
GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains
GDPx is a benchmark that covers the most valuable domains for frontier intelligence. It aims to be the world’s best benchmark for measuring general agentic capabilities.
We believe the first trillion dollars of AI revenue will be concentrated in sectors that can deploy frontier intelligence at high ROI, not broadly diffused throughout the whole economy. Building a general-purpose benchmark based on share of GDP or the SOC/O*NET occupational taxonomy is therefore a mistake. If your goal is to track the progress of frontier models, you should focus on the domains where they will be most impactful. — Read More
