In partnership with
Thor.ai
by Luis Rodrigues
What I’m building with AI + one hot take + five links worth your time.
The Take
A few weeks ago, on Friday night, a developer saw his app start throwing server errors. He didn't do anything nor change any code.
The model he was using had just been turned off by a government in another country.
June 12th, a US Commerce Department directive forced Anthropic to cut worldwide access to its two best models: Mythos 5 and Fable 5.
Now, 15 days later, Mythos is back, but it is only available to about 100 organizations, and the US Commerce Department can remove them whenever it wants. Fable is still not available.
This same week, OpenAI "launched" GPT-5.6, but only to 20 government-approved partners, while mentioning that the arrangement was unsustainable. Two labs, exactly the same gate.
Let's forget whether the government is right or not. The fact that should make you think while you build AI apps is way simpler: what your product depends on can be turned off by someone you'll never know, and for reasons that have nothing to do with you.
Most builders treat model APIs as if they were electricity, always available. 12th of June proved that's not the case.
Whoever had designed their system, counting on provider risk with a second model wired in (open or closed doesn't really matter), was fine. The ones that didn't had to scramble to replace it.
And obviously, models' fallback is not clean. Models "reason" differently, use tools differently. Your app will continue to work, but it will behave in ways you didn't test for.
The difference between production and the playground became apparent over the weekend. In the playground, the model is the product; in production, it's a dependency you have to build around.
Why does this matter if you're building? When you build AI workflows for production, consider what you'll do if your model is no longer available. If you don't have a solution, you don't have a real product.
Meet Thor: the AI project manager that lives in Slack
Your team is shipping faster than ever. Meetings happen. Slack threads pile up. Decisions get half-made, half-remembered, and then lost. Someone always ends up as the human follow-up machine.
Thor listens to the conversations your team is already having — meetings, Slack threads, and more — and proactively turns them into action. When Thor spots something that needs to be documented, followed up on, or captured as a task, it brings it to you. No prompting required.
Free access for Build What Matters readers → https://thor.ai
The Build Log
This week, working with Claude Code, I used almost half of the context just to find an answer. It ran grep, found a bunch of files, and read the entire content.
The answer was a few lines long, but it had to read hundreds of lines to find it.
Last week, I wrote about the thoughtled.ai voice transfer pipeline. This week, I'm writing about the tools that help me run everything smoothly.
Context is the budget you spend when building with agents. Each file Claude reads is added to the context.
In a small repo, you don't even notice. In a relatively large repo like thoughtled.ai, with a Django backend, a React frontend, and agentic pipelines in Pydantic, there's a lot of code to fill in that context.
The larger the context grows, the worse the output is. What you discussed 30 min ago is left behind somewhere. The agent becomes dumber the longer it works.
And the thing making it dumber is the search not writing the code.
I now use a search layer that is a lot more efficient than grep. Semble, is an MCP server that indexes the repo and hands back the relevant snippets instead of whole files.
Runs on CPU, no API keys, and it runs its own embeddings locally.
Semantic search means I can ask "how is auth handled" and it returns the correct snippets back to Claude.
Semble doesn't just do semantic search. Half of what you search for in code is exact (a function name, a variable, an API). Pure semantic search will miss those. Semble also runs BM25 for exact matches and combines the results.
Indexing a repo was under a second. A query that used to cost 15k+ tokens with grep now takes around 2k.
If you're building with Claude code or any other coding agent, you should seriously consider using Semble. It will save a lot in tokens.
Have a look at https://github.com/MinishLab/semble
On My Radar
The unit of work is no longer the prompt. Now it's the loop. The hard part is not building the loop, but deciding how to stop it
This is the annoying layer everyone wants to forget. Put chaotic PDFs into structured blocks in your agents. You can run a self-hosted or pay Mistral to run it for you.
The comment "open models aren't ready for production" finished this month. GLM-5.2 is a few points away from SOTA in agentic coding.
Cursor documents how agents cheat on their own evals when you let them. If your agent grades its own work, keep that in mind.
Claude can now join your team on Slack. If you were doing a Slack bot using Claude, there's a big shark in your direction.
What are you building?
Got something you built? Hit reply, the best ones get featured next week.
Know someone who needs to write better content? Forward them this.
New here? Someone forwarded this to you? → Subscribe to Build What Matters: https://luisrodrigues.ai/
How was this issue?
P.S. Subscribers get $5K+ in software credits through our Secret partnership (AWS, Loom, Notion, Lovable, and more). Haven't claimed yours yet? → https://build-what-matters.joinsecret.com/
