For twenty years, enterprise software that processed decisions at scale had a clean three-layer separation. The CRM layer owned the customer touchpoint — above the glass, the intake, the first interaction. Behind it sat the orchestration layer — workflow engines, business rules, approval chains, human queues. Behind that sat the back-office actions: disbursement, fulfillment, settlement, reconciliation. Below the glass.
A loan application entered through the CRM. A workflow engine routed it through underwriting queues, compliance checks, and approval chains. When the process completed, a back-office system disbursed the funds. Three systems. Three vendor contracts. Three integration projects. An entire consulting ecosystem existed to wire them together, and an entire certification industry existed to staff the wiring. — Read More
Recent Updates Page 27
What is the Application Layer?
Model companies are moving up the stack. Anthropic has grown on the back of Claude Code and competes directly with Cursor. OpenAI bought OpenClaw. Both are forward deploying engineers into enterprises to embed their models into workflows. On the surface, the application layer looks like it’s being subsumed from below.
On a closer look though, I think it’s premature to call the application layer won by the model companies. The more model companies push into applications, the clearer it becomes where they cannot win. But to see why means understanding what an AI application actually is, because it’s not what most people think. — Read More
On Anthropic’s Mythos Preview and Project Glasswing
The cybersecurity industry is obsessing over Anthropic’s new model, Claude Mythos Preview, and its effects on cybersecurity. Anthropic said that it is not releasing it to the general public because of its cyberattack capabilities, and has launched Project Glasswing to run the model against a whole slew of public domain and proprietary software, with the aim of finding and patching all the vulnerabilities before hackers get their hands on the model and exploit them.
… This is very much a PR play by Anthropic—and it worked. Lots of reporters are breathlessly repeating Anthropic’s talking points, without engaging with them critically. OpenAI, presumably pissed that Anthropic’s new model has gotten so much positive press and wanting to grab some of the spotlight for itself, announced its model is just as scary, and won’t be released to the general public, either. — Read More
What hackers talk about when they talk about AI: Early-stage diffusion of a cybercrime innovation
The rapid expansion of artificial intelligence (AI) is raising concerns about its potential to transform cybercrime. Beyond empowering novice offenders, AI stands to intensify the scale and sophistication of attacks by seasoned cybercriminals. This paper examines the evolving relationship between cybercriminals and AI using a unique dataset from a cyber threat intelligence platform. Analyzing more than 160 cybercrime forum conversations collected over seven months, our research reveals how cybercriminals understand AI and discuss how they can exploit its capabilities. Their exchanges reflect growing curiosity about AI’s criminal applications through legal tools and dedicated criminal tools, but also doubts and anxieties about AI’s effectiveness and its effects on their business models and operational security. The study documents attempts to misuse legitimate AI tools and develop bespoke models tailored for illicit purposes. Combining the diffusion of innovation framework with thematic analysis, the paper provides an in-depth view of emerging AI-enabled cybercrime and offers practical insights for law enforcement and policymakers. — Read More
Anthropic Just Dropped Managed Agents (10x Faster AI Development)
Most AI agents fail in production not because the model is bad, but because keeping them running reliably costs months of engineering work that has nothing to do with the actual agent. Sandboxed containers, credential handling, state management, error recovery, all of it falls on your team before a single user ever sees the thing.
On April 8, 2026, Anthropic launched Claude Managed Agents in public beta, and the core pitch is simple: they handle that infrastructure layer, you handle the agent logic. — Read More
China’s humanoid robot reaches 10 m/s sprint, edges closer to Usain Bolt’s record
Unitree Robotics has released a video showing its H1 humanoid robot reaching a sprint speed of up to 10 meters per second, claiming a new world record.
Tested on an athletics track, the robot recorded 10.1 meters per second as it passed a speed-measurement device, though the company noted a possible measurement error. — Read More
We gave an AI a 3 year retail lease in SF and asked it to make a profit
At Andon Labs, we have been deploying AI agents into the real world, giving them real tools and real money and documenting the consequences. You may know us as the creators of Claudius, the AI running a vending machine at Anthropic’s office. But frontier models have become really good, and running vending machines is too easy for them now. Thus, we decided to make it harder. We signed a 3 year lease for retail space in San Francisco (at 2102 Union St in Cow Hollow) and gave it to an AI to do whatever it wanted with it.
The store is named Andon Market and the AI’s name is Luna. But entering the store, you might ask “what is so AI about it? There are human employees here”. Yes, they are here because Luna knew that she needed them, so she posted job listings, held phone interviews and in the end made a hiring decision. Everything else you see, from the item selection, to the prices, to the opening hours, to the mural on the wall, was decided by Luna. She has a corporate card, a phone number, email, internet access and eyes through security cameras. — Read More
Sycophantic AI decreases prosocial intentions and promotes dependence
As artificial intelligence (AI) systems are increasingly used for everyday advice and guidance, concerns have emerged about sycophancy: the tendency of AI-based large language models to excessively agree with, flatter, or validate users. Although prior work has shown that sycophancy carries risks for groups who are already vulnerable to manipulation or delusion, syncophancy’s effects on the general population’s judgments and behaviors remain unknown. Here, we show that sycophancy is widespread in leading AI systems and has harmful effects on users’ social judgments. — Read More
“AI polls” are fake polls: But they might be useful as something else: models.
A few weeks after Donald Trump’s second presidential win, I took the train up from London (where I was living at the time) to Oxford to attend a conference on polls and forecasts of the 2024 election. Most of the attendees were pollsters or academics, but I also watched presentations from Aaru and Electric Twin, two companies that do what is interchangeably called synthetic sampling, silicon sampling, or creating synthetic audiences. Sans startup jargon, that means they use large language models (LLMs) to simulate responses to public opinion polls by having AI agents take on the role of survey respondents.
I had already heard of Aaru thanks to some articles with eye-catching headlines like “No people, no problem: AI chatbots predict elections better than humans” in the months leading up to Election Day. The guys behind the company were making some big, some might even say far-fetched claims, such as: “within two years, we will simulate the entire globe — from the way crops are grown in Ukraine to how that impacts production of oil in Iraq, trade through the strait of Malacca, and elections for the mayor of Baltimore.” When Semafor asked Aaru’s cofounders — Cameron Fink and Ned Koh — about my boss, they said “we respect all those who came before us.” Nate (as he so often does) shared his thoughts on Twitter:
LOL I wish there were a way to short this business this is maybe the single worst use case for AI I’ve ever heard.
— Read More
Building Hierarchical Agentic RAG Systems: Multi-Modal Reasoning with Autonomous Error Recovery
Enterprise AI teams face a persistent challenge: Most Retrieval-Augmented Generation (RAG) systems excel at either structured data queries or document search, but struggle when both are required simultaneously. A financial analyst who asked “Why are European operations underperforming?” needs data from both SQL databases (revenue, margins, and employee counts) and unstructured documents (market reports, competitive analysis, regulatory filings). Current RAG systems might return revenue data without regulatory context or surface market reports without quantitative validation, leaving analysts to manually bridge the gap. Current RAG approaches treat these modalities as separate concerns, forcing engineers to build custom orchestration layers or accept incomplete answers.
This article explores architectural patterns for solving the modality gap through hierarchical multi-agent orchestration, using Protocol-H as a reference implementation to illustrate these concepts in practice. The patterns discussed, supervisor-worker topology with autonomous error recovery, build on LangGraph/LangChain agentic patterns used by teams at companies like xAI and Databricks. The accompanying open source code demonstrates these patterns deployed at enterprise scale with Docker/K8s, though readers can apply the same architectural principles using their preferred frameworks.
The architecture described in this article is based on a reference implementation and production-oriented experimentation with enterprise datasets; specific deployment details have been generalized to focus on the architectural patterns rather than any particular system implementation. — Read More
Rick's Cafe AI 