AI-skeptics often ask (rightly) something along the lines of “If agents provide X times the productivity, why haven’t we seen X times the programming output?”
This is my attempt at documenting specific advances that are attributable to agents, and reasons why this argument isn’t as strong of a statement as it might first appear.
In general, my belief is that agents pre-Opus (November 2025) did not do much to accelerate development without significant effort creating harnesses to work around limited agent capabilities. Then it took a few months before non-AI-pilled (pre-AI-pilled?) developers actually got on-board.
The clock starts at November 2025…
Developers then needed time to actually recognize the improvement and to start using it themselves. For a lot of developers, they got a chance to play around with it around Christmas.
…or maybe even January 2026
From there, a substantial number of enterprises would put heavy barriers to agents running independently. No yolo-mode allowed. Agents could read the files system and write changes, but only commands in an allowlist would be run automatically. This is the phase where developers would joke about their job being to continually press enter.
It wasn’t until March/April that “Auto” approves were introduced where a separate LLM could validate commands on behalf of the user.
…or maybe even April…?
From there you had people learning how to effectively use agents in their own workflow.
So you might have people say for years that LLMs multiply programmer productivity. Those people might now claim that they saw the future and were there early. I don’t buy it. To me, the clock starts in Spring 2026.
Let’s get into specifics now.
Work-in-progress
2026-11-04 - Discord client moving to Rust - Discord was on Electron and seemed like it was going to continue this forever if agents didn’t allow them to reasonably swap to Rust. Not released, but I don’t think this would happen without agents