Eric Urban

LLMs are learning frontend from a web full of inaccessible patterns. That is a problem. In this sharp piece, Eric W. Bailey explains how to instruct LLMs to reference the ARIA Authoring Practices Guide carefully, extracting the useful guidance while avoiding blindly copying code examples. A thoughtful reminder that accessibility still requires human judgment. https://buff.ly/BP0j2If #Accessibility #AI #WebDevelopment #ARIA #InclusiveDesign #DigitalAccessibility #A11y

30

On Friday, Anthropic released a post describing their application's new "Research" capabilities along with how and why they used a multi-agent architecture. This seems like required reading for people implementing multi-agent systems. https://lnkd.in/g4__y4A8 "We’ve found that multi-agent systems excel at valuable tasks that involve heavy parallelization, information that exceeds single context windows, and interfacing with numerous complex tools."

5

The AI "industry" is not going to fall apart even if there is a significant correction in the marketplace. As with any technology, there will be companies that made the "right" choices and others that didn't, and ultimately, these will adapt to changes in prevailing technologies. I don't see OpenAI surviving, not necessarily because it's technology is increasingly inferior (it is, but that's never stopped a company before) but because it's business model is increasingly unsustainable compared to others. It will become the Lehman Brothers of this cycle, the sacrificial lamb used to bury all of the sins of the rest of the community, and it will likely take down a few others in its wake. Anthropic will survive primarily because the need for LLMs is still there, and will be a strong B player, but I agree with Mark here that it will likely be Google, Apple, and a somewhat scathed Microsoft that will remain standing at the end of all this. It may very well be the end of Oracle in any meaningful sense, but that decline was always only peripherally about AI.

14

1 Comment

The simultaneous emergence of Claude Cowork and Clawdbot (now rebranded as Moltbot) is a clear signal of where consumer AI is headed this year: genuine end-user productivity through semi-autonomous agents. The days of purely prompt-based, one-shot results are giving way to systems that can handle real, ongoing work based on high-level requests. These tools are starting simple, but their trajectory is unmistakable: they will likely evolve quickly into capable digital assistants running 24/7. If recent AI trends hold, most of us will soon dedicate at least one (if not several) always-on machines or cloud instances to handling our routine personal and professional chores. This shift could free up significant time and mental energy, allowing us to focus on higher-leverage activities with far greater returns. Security, however, remains the single biggest hurdle. Granting AI agents access to local files, email, calendars, or messaging apps introduces real risks such as prompt injection, data leakage, unintended deletions, or worse. While tools like Cowork use sandboxes and careful permissions, and self-hosted options like Moltbot keep data local, vulnerabilities persist and demand vigilance. For now, the wisest path is cautious experimentation: start small, isolate sensitive data, monitor closely, and contribute feedback to help shape safer, more reliable agents. These early steps are opening exciting new chapters for everyday users. AI, it appears, won't just answer questions, but will truly work alongside us.

36

3 Comments

Doing Things That Don’t Scale… at Scale One of the most fascinating shifts I’m seeing this year in AI adoption is the rise of the Forward Deployed Engineer (FDE) model. What used to be considered a niche, startup-y approach is now the dominant strategy for AI companies—and yes, even a $3T+ company like Microsoft is leaning into it. Why? Because AI agents are a new category. There’s no incumbent “product-market fit playbook” to follow. The product discovery has to happen inside the enterprise, shoulder-to-shoulder with customers. That’s where the FDE model shines: 🔹 Embedded at the front lines → Working directly with users, not just hearing about them secondhand. 🔹 High-value problem hunting → Pinpointing the 1 or 2 business-critical use cases that actually move the needle. 🔹 Step-function changes → Rebels/heretics in the best sense, imagining 3x–10x shifts, not incremental wins. 🔹 Echo team dynamics → Translating those scrappy “gravel road” solutions into paved superhighways for the platform team. Startups do this out of necessity. But the wild part? At Microsoft, we’re proving it works at global scale. That’s the paradox I love: Startups “do things that don’t scale.” Enterprises like Microsoft “do things that don’t scale—at scale.” As a Solutions Engineer, I’m convinced our role is evolving into this Echo Team DNA. We’re not just implementers—we’re co-discoverers, relationship builders, and multipliers of customer impact. If you’re curious, Bob McGrew (ex-Palantir, ex-OpenAI) breaks this down beautifully in the FDE Playbook → https://lnkd.in/gwH-WPTU The future of AI adoption isn’t about sitting back and waiting for product fit to arrive. It’s about embedding, listening, experimenting, and then scaling the wins into highways everyone can drive on. That’s how we make AI real. #AIAgents #FDE #ForwardDeployedEngineer #AIAdoption #SolutionsEngineering #TechInnovation

5

1 Comment

Microsoft released Visual Studio Code 1.112 on March 18, 2026, with a headline feature that gives Copilot CLI sessions a new Autopilot permission level -- one that auto-approves tool calls, auto-responds to questions, and works autonomously until a task is complete, with no confirmation dialogs required. Two additional tiers (Default Permissions and Bypass Approvals) give teams granular control over how much autonomy agents are granted. The release also adds sandboxing for locally running MCP servers on macOS and Linux, restricting their access to the file system and network, and a new /troubleshoot skill that reads agent debug logs and surfaces diagnostics directly in the Chat view. Agent context handling expands as well: agents can now read images and binary files natively, and pasting a copied symbol into the Chat input automatically creates a typed symbol reference. On the developer experience side, the integrated browser gains a full debug type (editor-browser), enabling breakpoints and variable inspection inside VS Code without switching to an external browser. A new monorepo-aware customization discovery setting rounds out the release, making it easier to share agent instructions across packages in large repositories without requiring the full repo to be opened as the workspace root.

2

Really great to see companies like Valo creating space for the community to come together, learn, and share what actually matters. Bringing admins, architects, and developers together to reflect on new releases, exchange perspectives, and learn from one another is how the ecosystem stays strong and grounded in real-world needs. Appreciate the leadership from Jari Salomaa and the Valo team for investing in community, conversation, and connection. This looks like a fantastic opportunity.

13

1 Comment