The Secret Is That You Already Knew

@systematicls, writing on X, in a thread titled "How To Be A World-Class Agentic Engineer":

You don't need the latest agentic harnesses, you don't need to install a million packages and you absolutely do not need to feel the need to read a million things to stay competitive. In fact, your enthusiasm is likely doing more harm than good.

There's a good post inside this post, trying to get out from under about 3,000 words of scaffolding.

The good post is this: most people working with coding agents have a context management problem, not a tooling problem. They install seventeen plugins, write a 26,000-line CLAUDE.md, pipe in memory systems from six sessions ago, and then wonder why the agent gets confused building a hangman game. Strip it back. Be surgical about what information the agent sees for a given task. Separate research from implementation. Specify what "done" looks like. That's useful advice, well-earned from real usage.

The sycophancy section is the standout. Asking an agent to "find a bug" biases it toward manufacturing one — because it wants to succeed at the task you gave it. The adversarial pattern of pitting a hyper-enthusiastic bug-finder against a sceptical disprover, then running a referee pass, is a clever design that treats the model's eagerness to please as a feature to exploit rather than a defect to complain about. More people should think this way.

But then there's the other post — the one that frames all of this as "the ascension piece you've been waiting for."

The core advice is "be precise about what you ask for." Which, yes. But telling someone struggling with agents to "just be precise about implementation" is like telling someone who can't cook to "just use better ingredients." The skill is the decomposition. The skill is knowing that you want JWT with bcrypt-12 and refresh token rotation before you sit down at the keyboard. If you already have that, you were probably fine. If you don't, this post hasn't actually given it to you.

The claim that you needn't bother staying current because frontier companies will absorb anything worthwhile is a comforting thought that happens to validate the author's minimalist setup. It's half true for tooling. It's not true at all for workflow patterns, domain-specific techniques, and the thousand small discoveries practitioners make that never get productised because they're not products — they're craft knowledge.

Once in awhile they do still get some things wrong, but this is now a nearly flawless exercise.

"Nearly flawless" and "once in a while they get things wrong" are doing an extraordinary amount of work in the same paragraph. For a post about precision, that's a suspiciously imprecise claim about the technique's reliability.

The irony is structural. A post warning against hype opens with "this is the ascension piece you've been waiting for." A post preaching minimalism runs to several thousand words. A post about not needing anyone's framework presents — a framework. The self-awareness is there, flickering, but it never quite lands.

Still: read the sections on context management and sycophancy. Skip the framing. The useful ratio is high if you already know what to look for — which, come to think of it, is exactly the problem the post claims to solve.