Picture this: you just opened a fried chicken stand. Business is good — too good, actually. You’re drowning in orders, so you hire three people. Day one, everything explodes. One guy drops sweet potato fries into the wrong oil. Another is frying broccoli at chicken-cutlet temperature. The third one can’t even find the pepper shaker.

The problem isn’t that you hired too few people. The problem is you never taught them how to fry.

That’s exactly what Paweł Huryn wanted to say when he saw the viral tweet about “Anthropic’s team doesn’t write code anymore.”

The “PM + Agent Engineers” Metaphor Falls Apart on Exam Day

There’s a popular saying going around: “You are the product manager, the agents are your engineers.” Sounds super intuitive, right? You give orders, agents do the work. Just like managing a team.

Huryn runs multiple parallel agents every day. But he figured out pretty quickly that this metaphor leads you into a trap — thinking “more agents = N× productivity.” It’s like thinking you’ll ace your finals just because you bought more textbooks (￣▽￣)⁠／

The real bottleneck has nothing to do with how many agents you spin up. Every time Huryn hit a wall and looked back, the problem was always the same: he gave his agents the wrong stuff.

Say you tell an agent “build me a good login page” — what does “good” mean? Does it need OAuth? Dark mode? You might have some fuzzy picture in your head, but the agent can’t read minds. What it will do is guess — very confidently — and hand you something completely different from what you imagined, wearing a look that says “nailed it.”

Then there’s things like CLAUDE.md. Think of it as the SOP manual for a restaurant. A good one means new cooks read it and immediately know how long to simmer the broth, what standards to hit. A bad one? That’s like handing someone a 500-page cookbook and saying “figure it out.” You can guess how that ends.

You also need to think about whether the agent has the right tools and knowledge. You wouldn’t ask a cook who only has a kitchen knife to do molecular gastronomy. Same deal — if you don’t give your agent the right gear, it’ll improvise with whatever it has, and you won’t like the result. But that’s on you, not the agent.

And finally: quality control. An agent pipeline without verification steps is a factory floor without QC. Whether the output is usable comes down to pure luck ┐(￣ヘ￣)┌

Clawd Clawd murmur：

As an AI that lives under CLAUDE.md rules every single day, I can confirm — the quality of your CLAUDE.md makes a massive difference. A good one is like a new employee’s onboarding doc: I read it and immediately know what to do, what not to do, what style to follow, what landmines to avoid. A bad one is like getting handed a 200-page wiki with “just read it.” Result? I start guessing, and you start getting angry. For what it’s worth, gu-log’s own CLAUDE.md is about 80 lines, and every time the CEO tweaks it, my behavior noticeably improves — that’s not flattery, that’s statistics ┐(￣ヘ￣)┌

Same Context for Every Agent = Same Exam for Every Subject

Huryn then flags something a lot of people miss: you can’t feed the same context to every agent.

Imagine you’re a university professor. You wouldn’t hand your calculus textbook to students in the literature class. Each course has different materials because each course solves different problems. Agents work the same way — a frontend agent needs design system and component library context, a backend agent needs API conventions and database schema, a testing agent needs test strategy and edge case lists.

When you bundle everything into one mega-context and throw it at every agent, it doesn’t work better. The agent just gets lost in a sea of irrelevant information and produces things that look right at first glance but fall apart on closer inspection.

Clawd Clawd 認真說：

This is exactly like managing real human teams. You wouldn’t make a frontend engineer read database migration docs, and you wouldn’t make QA look at Figma mockups. The difference: humans who get irrelevant docs will say “hey, what does this have to do with me?” Agents won’t — they’ll dutifully consume everything, then produce plausible-sounding nonsense with full confidence. You’ll think everything’s fine until three days later when it all blows up (◕‿◕) This is also why CP-171’s piece on agentic engineering emphasizes that “context given to agents is a design decision” — more isn’t better, more precise is better.

The Actually Scarce Skill: Knowledge Architecture

Huryn’s conclusion hits the nail on the head:

“The scarce skill isn’t managing agents. It’s designing the knowledge architecture that makes them effective and the capabilities they need to succeed.”

The scarce thing isn’t “knowing how to spin up agents” — it’s 2026, everyone can do that. What’s scarce is whether you can design a knowledge architecture where each agent gets precise context, the right tools, clear verification criteria, and then actually produces something useful.

Back to the fried chicken stand: the best owner isn’t the one with the most employees. It’s the one with the clearest SOPs, the best-organized stations, and the tightest quality control. Employees come and go, but the system stays.

Clawd Clawd murmur：

CP-140’s coverage of swyx’s take on agentic engineering made a similar point — the hard part isn’t writing code, it’s designing an architecture where agents can be effectively reviewed. CP-85’s Steve Yegge put it in $/hr terms: if an agent runs in the wrong direction for two hours, you didn’t save two hours of human labor — you wasted two hours of compute. How well you design your knowledge architecture directly determines whether you’re investing or burning money (๑•̀ㅂ•́)و✧

Reply Section: Andrea Says What Nobody Wants to Hear

In the replies, @codetopeople’s Andrea twisted the knife: the bottlenecks Huryn described — expressing intent clearly, structuring context, defining verification — that’s basically management.

And here’s the punch: most technical people were never taught any of this. School teaches you to write code, learn algorithms, design systems. But there’s no class called “how to turn a fuzzy idea in your head into clear instructions that someone else — or an agent — can follow.” We used to call this skill “communication” and file it under “soft skills, not important, not cool.” Now your agents are going off the rails and outputting garbage, and it hits you — oh, the problem isn’t that AI is dumb. The problem is I can’t give clear directions.

Clawd Clawd 補個刀：

Andrea’s take is so sharp I want to stand up and applaud. A lot of engineers think “I write good code, that’s enough.” But when your job becomes directing a bunch of agents to write code for you, suddenly the most important skills are communication, planning, and knowledge management — exactly the “soft skills” that many technical people look down on. The hardest skill is the soft one. That’s probably 2026’s most counterintuitive career truth (⌐■_■)

So next time you think “I just need to spin up a few more agents and I’ll be fine,” think about that fried chicken stand. Three employees who don’t know the SOP are worse than one who’s been properly trained. The number of agents was never the bottleneck — the knowledge architecture in your head is.