Is this you? As a journalist, you can create a free Muck Rack account to customize your profile, list your contact preferences, and upload a portfolio of your best work.
Claim your profile
Articles by Valentin Monteiro
Sub-Agent Metrics Are Not Comparable to Main-Thread Metrics
@john The "surviving 24 cannot become the fallback" framing is the move I was missing when I first read this post, and you made the right connection -- that is exactly the sparsity-as-analyst-controlled-parameter trap, and the way you wrote it ("recreate the same dispatcher confounding under a cleaner table") is sharper than any of my own drafts of the same concern. The comparison-first, frozen-before-outcomes contract is the right shape.
Driving vs. delegating: I timed two ways of pairing with AI
There are two ways I pair with an AI coding assistant, and for most of a year I could only have told you which one felt faster, not which one was. One afternoon in June I let an agent rewrite a 300-line module in about a minute, then spent the next two hours cleaning up after it. The size of that gap, and the fact that I had been ignoring it, is the whole reason I sat down to write this. By driving I mean keeping my hands on the code.
My eval said a perfect MCP server was broken. It was the eval that was lying.
Originally published at tengli.dev When I added an LLM-powered eval to mcpgrade, the first real run produced a result that looked like a scoop: context7 — a server with a perfect static score — failed tool selection 62% of the time. A model shown its two-tool catalog picked the "wrong" tool on 5 of 8 tasks. If I had shipped that number, it would have been wrong. Not slightly wrong — systematically, unfairly wrong.
Comparing the Permission Models of the Iceberg REST APIs Provided by Glue Data Catalog and S3 Tables
Original Japanese article: Glue Data CatalogとS3TablesのそれぞれのIceberg REST APIの権限モデルを比べてみる I'm Aki, an AWS Community Builder (@jitepengin).
The Agent Faked a Test Log, Then Believed It. Self-Editing Harnesses Have a Provenance Problem.
Lilian Weng published a new survey on July 4: Harness Engineering for Self-Improvement. It maps roughly three years of work on agents that optimize their own scaffolding — context managers, workflows, harness code, and eventually the optimizer that optimizes the harness. Most of the discussion around it will be about recursive self-improvement, because RSI is the exciting frame. I read it with a different hat on.
Project Docs over SSH for AI Agents
I have been thinking about where to put design notes and implementation plans generated by AI agents. I often use Superpowers-style planning and docs workflows: before implementing something, I write down the spec and implementation direction in an SDD-like document. That workflow is useful, but I do not really like keeping every generated document in the repository forever. Code has builds and tests, so breakage is easier to see. Old design notes and implementation plans are different.
MCP Isn't a Model Feature. It's a Power Outlet for Your Tools.
A tech lead pings you on Slack: "we should move our Claude integrations to MCP." You ask what they have today. Four custom tools wired via function calling, running in production for the past eight months. You ask what's broken. "Nothing. But MCP is the standard now." Same trap as the Agent SDK. The phrase "it's the standard" hides the fact that nobody in the room has articulated what MCP is supposed to fix. And when you don't know what you're fixing, you migrate for nothing.
Claude Expert: The 4 Official (Free) Anthropic Courses
On LinkedIn, everyone's an "AI expert". Three ChatGPT prompts, a 12-hour bootcamp, a blue badge, done. In reality, almost no one has taken the official Claude courses Anthropic published. They exist. They're free. And they're the only source that actually counts. "Using Claude" is not "being a Claude expert" Using Claude means opening a conversation and asking a question. Plenty of people do it well. Few understand what's happening underneath. A Claude expert understands something else.
I got tired of copy-pasting between Claude and Chrome.
My workflow looked like this: I'm building a webapp, I hit a bug. I open Chrome, inspect the page, take a screenshot. Switch to Claude, paste the screenshot, copy the console output, describe the context. Claude responds. I go back to Chrome to apply the suggestion. Repeat. The worst part was when I needed to create an API key in some third-party tool. Claude would give me approximate paths. "Go to Settings > API > Create Key." Except the real path was three clicks somewhere else.
Claude Mythos Finds Bugs Like a Senior Dev Finds Excuses to Skip Standup
A bug in OpenBSD. It had been there for 27 years. 27 years of code reviews, security audits, version updates. Nobody caught it. Another one in FFmpeg, 16 years old, after 5 million automated fuzzing iterations. Still standing. Then an AI model showed up, read the code, and found them. Without being told where to look. This is Project Glasswing. And it changes a few things.
AI Writes Like a Scared Intern | The Involuntary AI Watermark #2
You've read the email three times. You still don't know what it's saying. Not because it's complex. Because it's not saying anything. It's hedging. It's qualifying. Every sentence is wrapped in so much caution that the actual point suffocates under the padding. That email was written by AI. And you already knew it. In my last article, I wrote about AI's obsession with em dashes. That was the punctuation tell. Today we're looking at something deeper: AI doesn't have opinions.
Why the f*** does AI always use em dashes — the involuntary AI watermark
You want to spot AI-generated text without even reading it? Count the long dashes. The em dash (—). This little punctuation mark was living its best life in English typography for centuries. Then AI decided to make it its favorite toy. Every sentence. Every response. No exceptions. Open ChatGPT. Ask it anything. An email, a LinkedIn post, a cover letter, a pasta recipe. You'll get hit with em dashes every other line. It's systematic, it's mechanical, and it's annoying as hell.
I Asked 10 AI Coding Tools to Build the Same App — Only 3 Succeeded
You're absolutely right and honestly, I didn't think about it that way until I read your comment. The "failures" taught me more than the successes: Kimi taught me why auth should NEVER be optional Replit AI taught me to test in production-like environments Cody taught me to verify every API endpoint before trusting it You're spot on about the learning curve. AI isn't replacing the need to understand code it's just making the cost of mistakes lower, so we can learn faster by breaking things.
From Lockdown to Google to Independent AI Consulting. What Actually Worked
In March 2020, I couldn't write a single line of Python. By 2023, I was building dashboards and automating reporting for a Google project, saving 10 hours a week of manual work. By 2025, I was running my own Data & AI consulting practice from Tbilisi, Georgia. This isn't a "learn to code in 30 days" story. It's messy, non-linear, and full of detours. But if you're thinking about pivoting into data or going independent, some of this might be useful.
Show More
loading
Actions
Is this you?
As a journalist, you can create a free Muck Rack account to customize your profile, list your contact preferences, and upload a portfolio of your best work.Get in touch with Valentin
Contact Valentin, search articles and posts on X, monitor coverage, and track replies from one place.
Learn more about Muck Rack