Engineer Reports 5x Speedup From Coding Agents, and a Matching Drop in Skill

A developer writing for Towards Data Science says four months of running coding agents made them roughly five times faster at producing code, and roughly five times worse at the craft of writing it. The account, published under the headline 'AI Made Me 5x Faster. It Also Made Me 5x Worse at My Job,' describes one near miss that prompted the reassessment. The claim is anecdotal, but it names a tradeoff that vendors selling agentic development tools rarely put in the demo.

The setup will be familiar to anyone who has wired an agent into their editor. The author handed off implementation work, reviewed output, and watched pull requests land at a pace that would have been impossible solo. Four months in, the speed numbers held up. What did not hold up was the author's own ability to reason through the code without the agent in the loop, a decline they only noticed when the tooling failed or produced something subtly wrong.

That near miss is the pivot of the piece. An agent generated code that looked correct, passed whatever checks were in place, and carried a flaw that surfaced close to production. Catching it required exactly the kind of close reading the author had stopped doing, because the agent had been doing it. The post does not name the company or the specific system involved, which limits how much weight the incident can carry as evidence.

The question the author says almost nobody asks is what a developer is supposed to do while the agent writes. Reviewing diffs is not the same as designing a solution, and the gap between those two activities is where judgment gets built. If the agent handles the writing, the human keeps the approval step and loses the reps. Skills that go unused for a quarter do not wait around at full strength.

The productivity math around coding agents usually counts output: lines merged, tickets closed, time to first commit. It rarely counts what the engineer gave up to get there. Teams adopting these tools at scale have no standard way to measure whether their senior engineers are still senior, and no benchmark that would catch a slow erosion of debugging instinct before it shows up as an incident.

What to watch is whether anyone builds that measurement. Vendors have every incentive to publish speed benchmarks and none to publish skill-decay studies, so the data will likely come from engineering orgs running their own internal reviews. The open question is whether deliberate practice without the agent, scheduled the way teams schedule testing, can hold the line, or whether the speedup simply is not worth what it costs.