Sander Hoogendoorn

You already paid for Microsoft 365 Copilot licenses… But do you have one Copilot Agent that actually saves hours every week? If your company has Microsoft 365 Copilot and still hasn’t built a truly effective Copilot Agent, don’t start with “cool ideas”. Start with the right 5 questions (these will reveal the best first business use case): 1. Where do we lose the most time every week in repeatable work?(Searching, rewriting, summarizing, reporting, answering the same questions, preparing the same documents.) 2. Which workflow is “high frequency + low creativity”? If people do it daily/weekly and it follows a pattern, it’s a prime candidate for an Agent. 3. What information do we keep re-explaining across teams? Policies, procedures, product/service details, templates, project context, customer rules, approval steps. 4. What mistake costs us real money or risk when it happens? Wrong data, missing steps, inconsistent messaging, outdated templates, compliance gaps, slow approvals. 5. What can we standardize into a “single best answer” with clear sources? If you can define the “truth” (SharePoint, OneDrive, approved docs, SOPs), you can build an Agent that is reliable, not “just smart”. If you can answer these 5 questions, you can usually define your first Copilot Agent use case in 30–60 minutes. Then we build it properly: clear scope, trusted knowledge, measurable impact, adoption plan. If your company already has Copilot licenses and you want to create your first real business-value Agent, comment “AGENT” or DM me. #ai #agent #copilot #m365copilot #business #artificilintelligence

9

1 Comment

You shipped the feature. Three weeks later you shipped it again, because "the requirements changed." Nobody remembers the second build as thorough. They remember it as slow. That reputation cost is real, and it is almost never about your code. It is about five minutes of questions nobody taught you to ask before you wrote the first line. Here is the split. Every feature hides three kinds of questions. Only the first one gets asked out loud. The first kind: what does this do? A user hits submit, a record gets created, an email goes out. The PM hands you this one for free. Nobody skips it. The second kind: how well does this have to do it? Nobody hands you this one. The third kind nobody asks at all: which part of this is still being argued about? Say the feature is an inventory count on a product page. Who else reads that number besides the shopper? Is the warehouse dashboard reading the same field, or a stale copy somebody forgot to wire up? What happens the week a sale sends traffic to ten times its normal load, not the week you demoed it with three test orders? And which part of this, right now, is two stakeholders still disagreeing about in a thread you were never copied on? That last question is the one that causes the rework. The part still being fought over has not settled. Build it first and you are building on ground that is still moving. You will tear it out the moment it stops. Build that part last. Build the parts nobody is arguing about first. Those hold still. You do not prevent rework by typing faster. You prevent it by knowing which part is still moving before you commit to it. Ask those questions before you open the editor and you stop looking like the engineer who missed something. You start looking like the one who already knew where the ground was going to shift.

Yep... Whatever value you think sits in your data - even if well structured (which it rarely is) - the real value only emerges when malleable software acting on that data - gets operationalised - driving decisions and outcomes. We’re rightly allergic to vendors who oversold services and collected salaries without ever delivering operational impact. But swinging the other way—buying a pile of shallow GenAI tools just because they come in a shiny product wrapper is not any better... Interoperable data and tools are meaningless without the deployment that actually produces results on the ground / in the field...

4

🌐𝗧𝗵𝗲 "𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗪𝗲𝗯" 𝗶𝘀 𝗻𝗼 𝗹𝗼𝗻𝗴𝗲𝗿 𝗮 𝘁𝗵𝗲𝗼𝗿𝘆. 𝗜𝘁’𝘀 𝗮 𝗣𝗿𝗼𝘁𝗼𝗰𝗼𝗹." For years, we’ve forced AI agents to interact with the web like humans: squinting at screenshots, guessing which <div> is a button, and burning millions of tokens on messy HTML scraping. That era is ending. Google Chrome just introduced the early preview of WebMCP (Web Model Context Protocol). This isn't just an update; it’s the "𝘜𝘚𝘉-𝘊 𝘮𝘰𝘮𝘦𝘯𝘵" for AI-to-Web interactions. 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗰𝗵𝗮𝗻𝗴𝗲𝘀 𝘁𝗵𝗲 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆 𝗳𝗼𝗿 𝟮𝟬𝟮𝟲:   • 𝗙𝗿𝗼𝗺 𝗦𝗰𝗿𝗮𝗽𝗶𝗻𝗴 𝘁𝗼 𝗖𝗮𝗹𝗹𝗶𝗻𝗴: Instead of "best-guess" parsing, agents now invoke structured Tool Contracts. If your site has a search_products function, the agent calls it directly with precision.  • 𝗧𝗵𝗲 𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆 𝗚𝗮𝗶𝗻: By eliminating raw DOM dumps, we’re seeing a massive reduction in latency and token costs. Speed is the new moat.  • 𝗧𝗵𝗲 𝗗𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝘆 𝗦𝗵𝗶𝗳𝘁: "Tool descriptions" are becoming the new Meta Descriptions. SEO is evolving into Agent Engine Optimization (AEO). How well can an agent understand and use your site’s capabilities? 𝗧𝗵𝗲 𝗧𝗲𝗰𝗵𝗻𝗶𝗰𝗮𝗹 𝗕𝗿𝗶𝗱𝗴𝗲:  • 𝗗𝗲𝗰𝗹𝗮𝗿𝗮𝘁𝗶𝘃𝗲 𝗔𝗣𝗜: Progressive enhancement at its finest. Add toolname and tooldescription attributes to your existing HTML forms, and they instantly become agent-ready.  • 𝗜𝗺𝗽𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗔𝗣𝗜: For the heavy lifting. Use navigator.modelContext.registerTool() to expose complex JavaScript logic directly to the browser’s AI. 𝗧𝗵𝗲 𝗕𝗼𝘁𝘁𝗼𝗺 𝗟𝗶𝗻𝗲: If you are building web products today and aren't thinking about your "𝘈𝘨𝘦𝘯𝘵 𝘜𝘟," you are designing for a shrinking demographic. The web is being rewritten as a structured database for AI. #AIAgents #WebMCP #FutureOfWeb #ProductStrategy #GenerativeAI #GoogleChrome

14

The most useful thing I heard at DX's Amsterdam panel was this: AI at scale is not really a coding story. It is an operating model story. Leaders from Booking, TESCO, Adyen, and OLX kept coming back to the same point: code generation is only a small part of engineering work. Writing code may be 15 to 20 percent of the day, but the rest goes into documentation, security, observability, product discovery, reviews, coordination, and all the messy work around delivery. So if an AI strategy stays inside the IDE, it misses most of the opportunity. The second theme was that scale makes the hidden costs impossible to ignore. At TESCO, token costs for some engineers are already hitting 25 percent of salary. At OLX, a Figma-to-merge-request workflow reached 80 percent acceptance, but only after four to five months of foundational cleanup. In both cases, the lesson was the same: once AI moves beyond experimentation, it has to be treated like infrastructure. That means budgeting for it, governing it, and making sure ownership, documentation, and architecture are solid enough to support it. The third theme was measurement and context. If a company did not have a productivity baseline before AI, it now has a very weak way to prove impact. And at the same time, the codebase context problem is getting worse fast. MRs are getting bigger. Review depth is getting thinner. Fewer developers fully understand the systems they work in. That is why companies like Booking and OLX are investing in living documentation and repository context files. Not because it is nice to have, but because an agent is only as useful as the context it can trust at 3 a.m. That is probably the real divide in AI adoption. The winners will not be the teams with the most AI demos. They will be the teams with the cleanest foundations, the clearest ownership, and the best context across the whole software delivery lifecycle.

8

… and the winners of this year’s 💼 hackathon are George Hall, Alex Dernie and Leo Strijbos 🙌 They built self-improving code that used an agent to monitor infrastructure alerts. When a new alert comes into Slack, the agent starts investigating it and pushes a code change if it’s able to fix the issue. We can monitor the agent’s progress in Slack directly and message it if we want to get involved.. and the agents knows when to hand off challenging issues for a human to look into. All of this, built in 3 hours 💥 Briefcase processes millions of transactions a year, so building internal agents that enable us to serve this demand reliably is very useful!

22

1 Comment

Build production-ready browser-based AI agents! Notte is an all-in-one platform for building browser agents that are meant to run in production. Instead of stitching together tools, it gives you a purpose-built environment for browser automation. You get a proper IDE with a code editor, a live browser, and full execution logs. Agents have real identities for authentication, secure vaults for secrets, built-in proxies, and one-click deployment so prototypes can turn into reliable services. Here’s why this is a game changer: • Agent identities: Give each agent its own email and phone number so it can handle logins and 2FA autonomously. This turns browser agents into auditable, first-class actors instead of shared credentials or brittle scripts. • Automation studio: A single place to build and debug browser agents. Write code, watch the live browser as the agent runs, and inspect step-by-step logs when something fails. • Agent mode: Start with natural language and end with code you own. Describe a task, let the agent run it in a real browser, then convert that run into editable code you can refine and ship. • Demonstration mode: Click through a workflow once and let Notte turn it into a deterministic script. Useful when it’s faster to show the system what to do, while still ending with code you can control and extend. This is the kind of setup you want when browser agents move from experiments to real infrastructure. I’ve shared the link to the agent platform in the comments!

187

17 Comments