The Current State of AI Engineering

Part 1: TL;DR: When Agents Write Code Far Faster Than You Can Read It

The single most important software engineering skill today is reading comprehension.

I've said that multiple times over the past three months, and the data is starting to bear it out. A longitudinal study of 800 professional developers, tracking 151 million IDE telemetry events over two years, found that developers who adopt AI coding tools type significantly more code per month than those who don't. They also delete significantly more. Output volume goes up. So does the curation burden.¹

The part worth sitting with: 56.5% of those same developers reported that their time spent coding went down. The telemetry said the opposite. Developers using AI tools generated more than seven times the additional keystrokes per month compared to those who didn't. What people perceive and what they're actually doing have quietly diverged.

The generation problem, the thing the industry has spent years racing to solve, is largely solved. The gap between "I think I'm writing less" and "the code keeps multiplying" is where the risk is accumulating.

More than 1.1 million public repositories now import an LLM SDK, up 178% year over year. 80% of new developers on GitHub use AI tools within their first week.² The tools are embedded. The question of whether to use them is settled. The question of how to honestly evaluate what they produce is not.

We're generating code faster than we can honestly evaluate it.

This Isn't Just a Code Problem

It's a governance problem. Organizations are deploying agentic AI ahead of the policies and oversight structures needed to manage it. The gap between running a pilot and having mature governance around autonomous code generation is wide at most companies, and it's widening faster than most governance teams move.

And it's a labor conversation too. The economics of AI-assisted development are a CFO-level discussion now, not just an engineering one. That conversation is happening whether engineers are in the room or not.

What This Means for Engineers

The engineers who will be most valuable aren't the ones who write the most code. They're the ones who can read it fastest, evaluate it critically, and design systems that make generation reliable in the first place. The ones who can write a clear spec, build meaningful guardrails, and create an environment an agent can actually work inside.

Reading comprehension is the skill.

The job isn't writing code anymore. It's making sure the code that gets written is safe to ship.

Next up: the tool market. It's fracturing fast. There are four tools that actually matter right now, and the one winning enterprise RFPs isn't the same one winning over the engineers who'll write tomorrow's RFPs. There's also a new category starting to emerge that could shift the whole picture.

Sources

¹ Sergeyuk et al., "Evolving with AI: A Longitudinal Analysis of Developer Logs." arXiv:2601.10258, January 2026. Accepted to ICSE '26. Mixed-method study of 800 professional developers across two years, combining 151 million IDE telemetry events with a survey of 62 practitioners. https://arxiv.org/abs/2601.10258

² GitHub Octoverse 2025. Published October 2025. https://github.blog/news-insights/octoverse/octoverse-a-new-developer-joins-github-every-second-as-ai-leads-typescript-to-1/