AI's coding ability is outpacing our ability to wield it effectively. That's why all the SWE-bench score maxxing isn't syncing with the productivity metrics engineering leadership actually cares about. When Anthropic's team ships a product like Cowork in 10 days and another team can't move past a broken POC using the same models, the difference is that one team has closed the gap between capability and practice and the other hasn't.
That gap doesn't close overnight. It closes in levels. 8 of them. Most of you reading this are likely past the first few, and you should be eager to reach the next one because each subsequent level is a huge leap in output, and every improvement in model capability amplifies those gains further.
The other reason you should care is the multiplayer effect. Your output depends more than you'd think on the level of your teammates. Say you're a level 7 wizard, raising several solid PRs with your background agents while you sleep. If your repo requires a colleague's approval before merge, and that colleague is on level 2, still manually reviewing PRs, that stifles your throughput. So it is in your best interest to pull your team up.
From talking to several teams and individuals practicing AI-assisted coding, here's the progression of levels I've seen play out, imperfectly sequential:
The 8 Levels of Agentic Engineering
I'll address these two zippily, mostly for posterity. Skim freely.
Tab completion is where it started. GitHub Copilot kicked off the movement. Click tab, autocomplete code. Probably long forgotten by many and skipped entirely by new entrants to agentic engineering. It favored experienced devs who could adeptly skeleton their code before AI filled in the blanks.
AI-focused IDEs like Cursor changed the game by connecting chat to your codebase, making multi-file edits dramatically easier. But the ceiling was always context. The model could only help with what it could see, and annoyingly often, it was either not seeing the right context or seeing too much of the wrong context.
Most people at this level are also experimenting with plan mode in their coding agent of choice: translating a rough idea into a structured step-by-step plan for the LLM, iterating on that plan, and then triggering the implementation. It works well at this stage, and it's a reasonable way to maintain control. Though we'll see in later levels less of a dependence on plan mode.
Now the fun stuff. Buzz phrase of the year in 2025, context engineering became a thing when models got reliably good at following a reasonable number of instructions with just the right amount of context. Noisy context was just as bad as underspecified context, so the effort was in improving the information density of each token. "Every token needs to fight for its place in the prompt" was the mantra.
... continue reading