Lab

Always-on agents are systems whose future behavior depends on durable state accumulated across earlier interactions. We treat them as persistent-state systems: the operative system includes retrievable memories, but also task ledgers, permissions, credentials, commitments, provenance and audit records, shared state, trigger conditions, and externally committed effects linked to those records. The survey reads the literature through six diagnostic axes for each state item, authority, scope, mutability, provenance, recoverability, and actionability, and through a lifecycle in which state is written, validated, organized, retrieved, acted upon, updated, forgotten, audited, and sometimes rolled back. Across a 435-work coded corpus, treated as a scoped map rather than an exhaustive census, the literature concentrates more heavily on accumulating and retrieving state than on governing, recovering, or relinquishing it. We therefore introduce the Always-On Evaluation Protocol (AOEP-v0), a pilot evaluation contract that makes these governance requirements concrete by scoring state mutation and recovery obligations rather than answer quality alone. The resulting agenda connects always-on agents to databases, distributed systems, formal methods, capability security, and machine unlearning.

Memosa — the AI analyst for investment committees.

Claude is Anthropic's AI, built for problem solvers. Tackle complex challenges, analyze data, write code, and think through your hardest work.

You can automate hreflang tags using vector embeddings + cosine similarity - and it’s SO EASY! I presented this at Search 'n Stuff, and now turned it into a blog post with more instructions. As an example, I took pages from Semrush and compared a sample of hreflang tags vs the AI suggestions: 89% matching. It included pages in English, German, Spanish, French, Italian and Japanese. I'm sure there are edge cases, but having this *draft* done by AI can save so much time. Honestly, boring time we'd spend hours (days?) doing it manually. Just upload your crawl from Screaming Frog and get hreflangs done for you in seconds.

Google is re-inventing Chrome for AI Agents: $3.7T tech giant quietly launched an early preview of WebMCP in Chrome 146, which turns websites into APIs for AI agents 😳 The problem Google’s solving is simple: browser agents today kind of suck. They squint at screenshots. Parse messy DOM trees. Click buttons like confused interns. Slow. Fragile. Expensive. WebMCP flips the model 🤖 Instead of forcing agents to mimic humans, sites can now expose structured “tools” directly to them via navigator dot modelContext. Think: From screen-scraping → to API calls. Then, websites stop being pages. They become callable services for AI. Early tests show: → 89% fewer tokens → 97.9% task success rates → ~53% lower cost → Declarative APIs for simple forms → Imperative APIs for complex JS flows Booking a flight? Filing a support ticket? Adding to cart? No more pixel guessing. Just validated JSON schemas + direct execution. And this isn’t about a nicer browsing experience. It’s about shifting control. When agents can reliably act across sites, the default interface to the web becomes conversational. And no, SEO won’t disappear. But AEO - Agent Experience Optimization - will matter much more. The web was built for humans. The next version is being built for agents. P.S. check out how I built an AI operating system to run a startup with Claude 🤖: https://lnkd.in/dv7-ZRVc

Claude code is great for prototyping and building code for startups, but you still need a systems design expert and someone who understands security for enterprise use. The person using Claude Code still needs to understand exactly what it is building and what technical debt is will need to be resolved, how the current state of the system fits in with the broader scope of a companies ecosystem. I’m absolutely not impressed by Claude Code’s ability to understand security best practices. Ask it to build a secure system and watch it go into a doom loop where is not absolutely sure what it’s built is secure and it loses confidence and tries refactor after refactor. The big insight is that when it’s done looping, neither you or Claude code is confident that what you have built doesn’t contain a critical security flaw that an experienced system architect would have caught.

Do you need entity data in your content workflow just as much as I do? Well you're in luck! This mcp server is for Google's free public Knowledge Graph Search API - search for structured entity information about people, places, organizations, and concepts, knit the data into your stuff 🧶 New from Houtini. You're welcome.

Confused by LangChain, LangGraph, and Deep Agents SDK? Here's the mental model that finally made it click 🧠 After diving deep into LangChain's official docs, I realized there's a clean 3-layer architecture that explains how all their products fit together: 🔄 Runtime (LangGraph) - The foundation Low-level orchestration for production agents. Durable execution, streaming, human-in-the-loop, state persistence. Use when you need fine-grained control. 🔝 Framework (LangChain) - Easy abstractions Higher-level abstractions built ON TOP of LangGraph. Structured content blocks, agent loops, middleware. Use when you want to quickly build agents without complex orchestration. 🧰 Harness (Deep Agents SDK) - Batteries included Opinionated framework also built on LangGraph. Planning capabilities, subagents, file systems, token management. Use for complex, multi-step tasks. The key insight? Both LangChain 1.0 and Deep Agents SDK are built on LangGraph. Understanding this hierarchy helps you pick the right tool: → Quick prototyping? LangChain → Production deployment? LangGraph → Complex long-running tasks? Deep Agents SDK And yes, you can combine them! Source: https://lnkd.in/gGy-Ghdg What's your go-to for building agents? hashtag#AI hashtag#LLM hashtag#AgentEngineering hashtag#LangChain hashtag#LangGraph hashtag#AIEngineering