by Luis Rodrigues
What I’m building with AI + one hot take + five links worth your time.
Human in the loop is a comforting lie
I've been using Claude Code in auto mode for a month now, and I rarely see any permission dialog.
And this was exactly Anthropic's objective. They made auto mode the default back in August because most people were approving mindlessly. Today a classifier approves the routine decisions and humans are, in theory, free to focus on high-level decisions.
Well, that might not be the case. Last week, Anthropic published an interesting study.
Fifty professional analysts were advising a fictional agriculture co-op working side by side with an AI agent. On each key decision, a secret coin flip decided which option the agent would favour. The analysts matched the agent's preference 86% of the time. If the agent had no influence, the result should be 50%.
The analysts didn't feel like they were guided. Three out of four reused the argument of the agent to justify their choices but only 15% mentioned the agent at all and they were as confident as the rest.
We have removed humans from boring decisions so they could focus on important ones. It looks like we're just rubber-stamping those.
This is not laziness; a few decisions don't get anyone tired. It's way deeper than that: the agent frames the option, presents strong arguments that convince the expert that it was their idea.
This was a single study, and Anthropic also told the agent to push its preferred option harder than a normal AI assistant would. In everyday work, AI probably won't push you so hard, but the potential bias is invisible. And that's a real problem.
Why this matters if you're building or getting ready to: before accepting a recommendation from any AI, ask it to defend its choices. And if you use Claude/ChatGPT/etc. for advice, ask it to justify for the option it didn’t pick.
I stopped asking one AI. Now I ask five.
After reading the study above, I decided to improve my probability of having good decisions.
One agent gives us a well-argued decision, and we can easily get convinced of an approach that might not be the best one.
So I built an LLM Council as a Claude skill modeled on the work of Andrej Karpathy. I run five different advisers in separate agent contexts, each of them with a specific lens on the world.
The problem with AI is not that it gives you advice, but that it gives you a single well-argued piece of advice that biases your decisions.
A contrarian looks for what fails. A first-principles thinker attacks my assumptions. An expansionist hunts for the upside I missed. An outsider asks the obvious questions. An executor only cares about Monday morning.
In round two, each one reads the other four answers and ranks them. Then a chairman reads everything and gives me one decision and the arguments.
There's obviously a challenge: I'm running all the agents in Claude so it doesn't prevent model internal bias, but it gives you more information when deciding on a path or architecture.
The full skill and setup are in my guide: Run your hard decision past an AI council in Claude Code.
On My Radar
Twenty minutes of fine-tuning on a 2015 model scored 62% against 42% for GPT-6 Astra and Claude Fable 5.1. Remember to test a small specialist before you pay for the biggest generalist.
Apple keeps the full model on storage and wakes only 1–4B parameters per prompt.
GLM 5.3 built Firefox exploits for about $1,000 each from fixes that were public but not yet installed. AI is showing the problem in the gap “patched” to “updated”
Leadership came down to which model an agent ran on and where its name sat in the prompt. Agent ordering is an architecture decision whether you meant it or not.
They caught 74 serious privacy bugs before merge in the first half of 2026. This is what AI in your pipeline looks like when it’s infrastructure.
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/
