The age of delegation

When you work, the why, the what, and the how all matter. Why do it, what to do, how to do it. From high level to low level, from planning to execution. But the three layers don't look quite equal.

Of course, the purpose or essence of the work (why) isn't necessarily superior or prior to its object (what), nor the object to its procedure (how). Constraints on method sometimes change the implementation goal, and a method that only someone can wield sometimes becomes the essence itself. It's a see-sawing relationship.

Still, by and large, why > what > how. You only do anything because there's a reason, and only once there's a goal do you rack your brain over how to achieve it.

2023 brought a flood of LLM-based toys. Even the rudimentary coding agents, whatever grand ambitions they carried, were just toys back then. There was a time when we struggled alongside ChatGPT, or vscode Copilot, to squeeze out just a few lines of code.

By 2024, hints were drifting in from everywhere that agentic coding might soon be possible. Cursor, which had watched vscode Copilot's lukewarm updates and decided "fine, I'll do it myself," started becoming genuinely usable in 2024. Windsurf's Cascade, deployed that November, gave early adopters quite a shock, as I remember. Agentic multi-file editing; my goodness.

In February 2025, the Claude Code CLI shipped. The community rippled. Good ideas are hard to think of, but once you see one, you slap your knee. Coding agent CLIs sprouted everywhere.

From mid-2024, when I began using coding agents in earnest, until mid-2025, I mostly flailed about trying to dictate the what and the how together. Tell these silly coding agents only briefly what to build, and they would drag in some absurd implementation approach and trash the codebase.

I became a master prompt-whittler, typing out, word by careful word, "this goes like this, that goes like that; use this pattern more than that approach," cramming the best practices I knew into their tiny context windows. They caused accidents anyway, and I stubbornly kept at agentic coding, simply grateful to be a 0.5x developer, never mind 10x.

That they would one day replace me cleanly, without a trace: I hadn't doubted that future for a single moment since witnessing ChatGPT's debut. But the future is the future, and the dumb AI agent right in front of me was simply deplorable.

Three months ago, Opus 4.5 was released, and at last my long fight with the how came to a close. I no longer write the how in my prompts. The agent has finally reached the stage of being entrusted with the what.

Delegating properly still takes care. What I can trust them with is still only the most basic layer of the what. The layers above? A matter of time.

The direction AI progress points to is so clear that I often catch myself lost in imagining it. Soon, the delegation of the what will be complete. Only the why will remain. And then it will be why's turn.