by Luis Rodrigues
What I’m building with AI + one hot take + five links worth your time.
Muse, Meta's Trojan Horse?
This week I heard Ben Hatton spending most of the time in a BBC panel defending AI agents. He's an analyst at CCS Insight and clearly the optimist in the room.
When the other members of the panel said agents are dangerous, he reminded everyone that there were people before who were afraid of railways
Then, interestingly, when the presenter asked if Muse, Meta's new personal AI, was useful or a Trojan horse, Hatton replied it was a Trojan horse.
Agents, in general, are going to be useful, but for this one, the panel voted three to zero that it's a Trojan horse.
The man who believes in agents doesn't believe in Meta.
Now let's look at what Meta launched in more detail. Each user has an isolated virtual machine and a separate security agent, Sentinel, that checks every action before it can access the internet. The agent can't access the passwords, which are protected in a vault, and Meta says it won't use the data for advertising.
Meta even delayed the release to harden the product's security. Serious engineering, but almost nothing was discussed in the almost half-hour BBC panel. Nobody debated the architecture because they were all focused on Meta's 20+ years.
Why this matters if you're building: with agents, trust is the product. Your history will be judged before your fancy architecture because agents have so much more access than a "normal" application.
I Transformed APIs into MCP tools
Thoughtled.ai has its own agent; our customers talk with it, and it helps them create LinkedIn posts. This week I opened another door to access the system.
I built an MCP server so people can use their favorite assistant to create content. All the guides I keep reading say the same: don't wrap the APIs. Webflow did a wrap; the agents didn't perform, and they had to rebuild MCP tools.
I still wrapped 16 of my 20 APIs, which made my life a lot easier for security and access control. A study I read in May found that, of 7,973 live remote MCP servers, about 40% exposed tools without any authentication.
So the MCP server uses the same APIs as our React app. The same login with the same permissions.
Two doors to access the product, but the same set of rules for both.
Your frontend API is normally built for intent, assuming you built it around the jobs your users need done. So it should be a good place to start to implement your first MCP.
Thoughtled has 20 APIs; 16 of them became MCP tools. The other 4 were too granular, so I didn't expose them. So Webflow was right for 20% of the product.
If you want to create a new MCP interface, check which API calls your frontend is doing that are real user actions. You can wrap all of those into an MCP tool. Don't forget to implement OAuth so you can pass identity back to your backend, and the access rules remain the same as the API.
On My Radar
A third of early users had already wired Muse into a business account, which tells you agents are reaching the people who never open a terminal.
Near-flagship work at $2 in and $10 out makes the mid-tier the production default and the flagship the exception you have to justify. Vercel's State of Agent Skills
A million skills in seven months, and most teach agents what everyone already knows, so the one worth writing is the one only your business could.
No verdict until late October, but the method already beats every "is it dumber today" thread.
Your MCP server can now push updates to ChatGPT instead of waiting to be asked, which moves MCP from tools you call to products that tell you things.
Launching my Instagram
I’ll publish daily guides on AI tools and processes. No hype.
Thanks for reading.
See you next week. Go build something.
Luis
Know someone who'd enjoy this? Forward it.
New here? Someone forwarded this to you? → Subscribe to Build What Matters: https://luisrodrigues.ai/
