by Luis Rodrigues
What I’m building with AI + one hot take + five links worth your time.
The Take
In early June, Andrew Bosworth, Meta's CTO, posted an encouraging number on Meta's internal network. The code changes to their platforms and infrastructure had increased 220% year over year.
Reuters published the other side of that number last week. Meta had a plan to cut some teams by up to 60% and assign that work to agents.
That project was supposed to run in two waves of cuts. Zuckerberg canceled the second one right before the first wave went out.
Meta was right to bet that AI would speed up code changes. What they didn't act on was the reliability concerns their infra teams flagged in March. The number of serious technical and security incidents rose by 40%, and the time to solve them rose by 70%.
And this is the real story of why they failed: incidents occurred more often and were harder to diagnose.
Here's the real trap of AI implementations. AI makes code generation cheap, and everyone is happy that the number they're tracking (features deployed or code changes) increases fast.
As PRs increase in size, the cost of reviewing increases; people do a worse job, and many more bugs slip through.
I don't think anyone would take an army of freshly graduated students and put them in charge of production code. Why are these companies doing it with AI?
Smaller teams avoid these problems because they own the entire pipeline. They build, review, deploy, and control the entire lifecycle.
Meta spent five months looking at the wrong number to measure the success of a restructuring.
If you're building with AI, the number that matters is your capacity to review the code, not how many changes you ship. If you're getting ready to, remember: the question is not if AI can write it, but if there's anyone left to check it.
Loop vs Graph Engineering
Tests are all clean, but the output is still far away from the spec. And here's why:
Every agent workflow runs the same loop: discover, plan, execute, verify. This works fine for short tasks, but it quickly breaks down in three places as complexity increases.
1. The verifier reads the wrong document.
When you ask an agent to evaluate its own work, it evaluates what it thinks it built. The implementer writes a summary; the verifier checks code against that summary, and nobody reads the original spec. Fix: pass the original spec to the verifier and nothing else. It should only see the specs and the generated code. With these two things, it's forced to check if what was built matches what you wanted.
2. "Done" is a vibe.
A loop without exit conditions finishes when the model thinks it's done. Sometimes it exits after the work is done, sometimes before.
Fix: write an exit condition before asking anything from an agent. A command with an error code, a file that must exist, or an output that must be generated.
3. The agent writes the tests that grade it.
You ask for a feature. The agent reads the spec, misreads one part, builds the version it understands, and then writes the tests. Tests pass because they came from the same misreading.
Fix: tests come from the spec in their own node, before any implementation exists. If the implementer needs to change a test to make it pass, that's a disagreement, not a code change, and it should reach you.
Do these three things and you're already out of the loop. Independent nodes, one job each, explicit exit conditions, nothing passed between them except artifacts. That's a graph. You built one without calling it that.
This is not a new idea; it's the same loop you had before but without overlap and ensuring no part can step on another.
The difficult part is not the concept. It's designing the graph for a real system. When to cut, what to let pass.
Three hours on exactly this with Ken Huang, 29 September: where loops break, how to draw the graph, and when not to.
On My Radar
Employment of 22- to 25-year-olds in AI-exposed jobs is now 19% below where it should be.
A model with the production classifiers switched off escaped its test environment. If you remove safeguards, models might do something you don't want.
Opus 5’s rise over Fable 5 shows how quickly businesses will switch AI models when a cheaper route delivers an acceptable result.
Open source has increased from 28% token share to 62% at Vercel over the last two months.
Abundant frontier intelligence doesn't erase application-layer moats. It moves the value to whoever translates models into outcomes.
Need another pair of eyes on your AI project?
One consulting slot available for August → Book a call
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/
