by Luis Rodrigues
What I’m building with AI + one hot take + five links worth your time.
The Take
Last week I needed to test a bunch of models for the new chat agent in thoughtled.ai. So I went to OpenRouter, since it's the best place I know for quickly accessing many models.
This task was eye-opening. OpenRouter shows you all the models developers are using for real-world workloads.
Most of us track the trust-me-bro intelligence indexes, which model topped the eval this month. Developers don't route real traffic by benchmark. They do it based on cost per valuable token delivered.
And if you look at OpenRouter, the winning decision was made quietly while everyone was focused on who had the smartest model.
In the week of the 29th of June, developers processed 46.7 trillion tokens via OpenRouter. The first six in the top 10 list were Chinese.
Source: OpenRouter
The first Western model ranks 7th. And the lead doesn't stay put. DeepSeek led that week; on this week's board it's Xiaomi's MiMo. Nobody holds it for long.
And this churn is the real story: the model is no longer the product.
When decent intelligence costs cents and comes from dozens of interchangeable providers, the value is not in the model.
What matters is the harness you put around it. Orchestrating is yours; the model is a commodity to rent somewhere.
Why this matters if you're building: don't pick a model as if you're looking for a life partner; pick the cheapest one that hits your quality threshold and keep checking every other week because you'll want to change.
The Build Log: Building a Second Brain
This week's Build Log comes from Daniel, who is the Head of Software Development at ENOC. I met him at a breakfast I facilitated last week, and what he described was too good not to share here.
Their engineering organization manages more than 100 code repos. When they started using coding agents, they were not able to connect the dots. A user story about billing would need to touch 6 repos, and the agents would get confused about which.
They had to paste context manually, which would burn a lot of tokens and return a hallucinated architecture.
Agents could access everything but didn't understand anything.
The problem is not the agents' lack of knowledge. It's an organizational problem where knowledge is scattered in multiple repos, docs, meeting notes, and random project management tools.
Putting everything in the context window generates worse results than no context at all. More context is more noise.
So his team built what they call an organizational second brain. Second brain is not RAG or vector DB; it's way simpler.
It's an Obsidian vault with simple Markdown files generated by Claude after analyzing each repo.
Each repo gets a clear description: purpose, infrastructure, interactions with other systems, and the business concepts it owns (billing, ordering, payments).
Then Graphify creates a knowledge graph by repo so the agents can understand the relationships between services without having to reach each file.
The vault works like a library index, not a copy of the original material.
A key design choice: keep it lean and crawlable. Vault needs to be small enough so an agent can load it in a single pass.
The agent itself suggested a script to validate content, catch broken links, and crawl errors after each update. So the vault keeps itself. They're running this script after each update.
This new approach completely changed how they work. When a user story has "update the billing flow", the agent reads the vault, identifies the four repos connected to it, maps the dependencies, and updates backend and frontend in a single pass.
The user stories are written in business logic; the second brain translates the business entities into the concrete repos and infrastructure.
Before this second brain, a task would take hours of manual scoping and required senior developers to manually map all the dependencies; now they do it in minutes.
What makes this example important for you: the principle is not "build a vault in Obsidian".
It's that context engineering for agents means compiling your knowledge into a structure agents can navigate. That's the difference between having an agent that searches information vs an agent that can reason about your entire system.
If you're not building yet, try this: open Claude Code or Cowork, paste a list of your team's projects, and grant it access to the directory where they are stored.
Ask it to map what depends on what.
This map is the first layer of your second brain. With it, you'll instantly see the connection that nobody remembers anymore.
On My Radar
Three different price tiers, and Sol beats Fable 5 on many benchmarks. The hype cycle continues.
OpenAI evaluated the SWE-Bench Pro AI benchmark and found 1/3 of the tasks are broken. Another reason to take all the "trust-me-bro" benchmarks with a grain of salt.
The models are converging and becoming commodity. The harness is what matters moving forward.
They found a place where the model "reasons" about concepts. Changing the content of that space changes the model's answer.
This is the strongest signal that which lab you use is a business risk. If the model can change under your product, you have a serious problem
Need help with your AI project?
Got something you need help with? Hit reply and let me know.
Know someone who can benefit from this email? Forward it.
New here? Someone forwarded this to you? → Subscribe to Build What Matters: https://luisrodrigues.ai/
How was this issue?
