Andrej Karpathy went from 80% manual coding to 80% agent-assisted in a single month last December. His takeaway: the intelligence is ahead of the infrastructure.
I've been thinking about this a lot. We're living through two shifts at the same time, and I don't think we're talking about them together enough.
1️⃣ The first is who writes the code. Coding agents like Claude Code are now responsible for somewhere between 16-23% of GitHub contributions, and that number is climbing fast.
2️⃣ The second is what the code does. Modern applications aren't purely code anymore. Behavior emerges from the interaction of code, model weights, and prompts — and it varies at runtime in ways we've never had to account for.
Here are my thoughts about what I think comes next: why traces are becoming the documentation layer for AI applications, why coding agents need their own observability stack, and what it looks like when the feedback loop actually closes — Agents that can instrument code, query their own telemetry, run evaluations, and submit changes with evidence.
This is the problem space we're building for. If you want to go deeper, I'll be demoing and having a lively discussion about this with Hamel Hussain in a free webinar where we'll get into the nuts and bolts — automating evals with Claude Code and Arize Phoenix. Link to register in the thread.
https://arize.com/blog/closing-the-loop-coding-agents-telemetry-and-the-path-to-self-improving-software/
