“Coding” Solved: What Happens to a Company? 7 Strange Things Happening Inside Anthropic That Could Be Your Script in 2-3 Years

Recently I finished watching that Lenny interview with Fiona Fung, Engineering Lead for Claude Code and Cowork at Anthropic. My spine went cold.

Not because of some jaw-dropping new technology revealed. On the contrary — she barely talked about how powerful the model is. She talked about all these trivial, almost messy internal changes: management disrupted, six-month roadmaps gathering dust, hiring standards shifting, engineers starting to feel lonely. Sounds like mundane company gossip.

But if you string these fragments together, you realize this isn’t “inside baseball at a hot company.” It’s a postcard from the future. Anthropic is special not because its people are smarter than yours — it’s because it hit the wall first, before anyone else: When output becomes free, what happens to a company? Its current frantic, scrambling state is likely what your company will look like in two or three years. So these 7 strange things are worth unpacking one by one.

Let me hand you the key first so you don’t get lost in the 7 items: Bottlenecks never disappear — they just move. When you make one step free, scarcity instantly rushes to the step next door. Coding got cracked open, so all the pressure, value, and job security on the assembly line whooshed straight to the next stage. Everything happening inside Anthropic is essentially the company frantically rebuilding its structure around that newly relocated bottleneck.

The Bottleneck Moved: From “Can It Be Written?” to “Can It Be Trusted?”

Fiona’s first take: coding is solved. Code is no longer the bottleneck; output has exploded 8x. But she immediately followed up with an even more important point: the bottleneck didn’t disappear — it moved to verification.

This is the most overlooked, yet most lethal asymmetry of the AI era: Generation got cheap; verification did not. You can make one person write 10x faster, but you cannot make that person “carefully read through and guarantee it’s correct” 10x faster. The machine frantically creates; the human vouches for it. And vouching is inherently slow.

So “8x output” sounds like a godsend, but the truth may be a trap: If humans still need to gatekeep, and gatekeeping is slow, then 8x output doesn’t make the system 8x faster — it just creates an 8x longer queue in front of the one person who can actually approve it. You didn’t speed up the work; you just made the line longer. That old, dusty Amdahl’s Law comes roaring back: you optimized the fast part (coding) like crazy, but didn’t touch the slow part (judgment), so the overall system didn’t speed up at all. The subtle panic in Fiona’s team? That’s exactly what it is.

Another signal: now designers, PMs, legal, even finance folks are starting to code. Many read this as the victory of “everyone becomes an engineer.” That’s backwards. It’s actually a sign that “writing code” has been thoroughly commoditized — something anyone can do. Once everyone can do something, it’s no longer a moat. The real moat has moved: who can judge whether this mountain of output is correct and worth shipping?

The Things Dying Are All “Prosthetics of the Scarcity Era”

This is the sentence I most want you to remember.

Over the past few decades, all the things we took for granted in companies — middle management, six-month roadmaps, grind culture, layers of approval — didn’t fall from heaven. They were all prosthetics* bolted on to deal with the same old disease: output was slow, expensive, and scarce.

Now that output is no longer scarce, these prosthetics collectively lose their reason to exist and start falling off one by one. The seemingly random “changes” in Fiona’s mouth, when translated, describe a company publicly molting — shedding layer after layer of old skin. Let’s look at exactly which layers are shedding.

The Dead Middle Layer: It Was Never “Medium Skill” — It Was “Glue”

Fiona says hiring now only has two categories: people with product intuition (“dreamers”) and strong systems experts who can chew through tough problems. The layer in between? No longer in demand.

Many interpret this as “AI replaced mid-level people.” That’s too shallow. The real function of the middle layer was never “medium skill” — it was translation and coordination. Turning a vague idea from the boss into actionable work for engineers. Aligning one team’s progress with another. Shuttling information between gaps, smoothing friction.

And AI happens to be a 24/7, zero-cost translation and coordination machine. So the middle layer is eliminated not because it was “too weak,” but because its entire reason for existence — the communication gap — got filled for free by AI. It was glue. And glue is now free. Why do the two ends survive? Because they’re two things AI can’t yet originate: “What should we build?” (taste, intuition, vision) and “How do we crack the hardest technical nut?” (truly frontier systems). The head and the tail are human; the middle was glue. Once the glue dissolves, the two ends snap directly together.

The Dead Roadmap: Long-Term Plans Were a Byproduct of Slow Output

Fiona says when she first joined, she ambitiously wanted to create a beautiful six-month roadmap. After three months, she realized no one was looking at it. Now it’s an Excel sheet listing monthly priorities, checked weekly.

Why did the six-month roadmap die? Because “long-term planning” itself was a response to slow output. Since building things used to be slow and expensive, you had to line up and lock down that scarce capacity half a year in advance — otherwise, coordination was impossible. Plans were a crutch for slowness.

But now building is so fast that a direction set in the morning can be overturned by a new capability that pops up in the afternoon. Under this rhythm, spending two weeks polishing a beautiful six-month plan isn’t rigor — it’s waste. By the time you finish it, it’s already expired. It’s like buying yesterday’s newspaper. So-called JIT (just-in-time) monthly planning isn’t laziness; it’s finally realizing that in a world where change outpaces planning, over-planning is an expensive form of self-comfort.

The Dead “Cult of Effort”: Burning Tokens Is the New “Lines of Code”

Here’s another particularly interesting shift. In early days at Anthropic, people competed over who burned the most tokens, who spawned the most agents — as if that proved competence. Then the collective lightbulb went off: that’s as hollow and stupid as comparing lines of code written. So the wind shifted — toward ROI and results, not usage.

This reversal is worth savoring because it punctures a deeply ingrained human flaw: we can’t stop measuring effort. In the past, effort and output were roughly proportional — more code written meant more stuff built. So just watch effort. But with AI, that ratio collapses entirely. One person can burn massive tokens and spawn a row of agents, producing nothing valuable. Another can use a single precise instruction and get the job done. When effort and outcome decouple, staring at effort is like looking for a sword where you dropped it in a moving river.

That’s also why doing this right is so hard. Watching results is much harder than watching usage. How many tokens burned is a straightforward, brainless number. “Does this thing actually have value?” — that’s a judgment call that requires thinking and taking responsibility. Humans naturally want to hide behind that easy-to-measure vanity metric. Watching usage is lazy; watching results is real work.

The Only Thing That Didn’t Get Cheaper: Human Attention

By now you can probably see it: every new rule that grew inside Anthropic is doing the same thing in its bones — rationing human attention. Because in an era where everything is getting cheap, the only thing that hasn’t dropped in price is a person’s ability to “carefully focus on the right thing.”

Her much-discussed “bad vs. sad” framework is a perfect example. “Bad” means irrecoverable damage (CLI crash, work lost); “sad” means recoverable annoyance (UI flicker). On the surface, it’s quality grading. At its core, it’s attention rationing: you can’t verify everything anymore (remember the verification famine?), so you must triage — save lives first, then soothe itches. And adding “sad piled up becomes bad” draws a warning line for itches too, so minor annoyances don’t accumulate into major disasters.

Look back at the other rules — all rationing: JIT planning means “don’t waste attention on a future that will change anyway”; not tracking tokens means “don’t waste attention on fake effort”; cutting the middle layer means “don’t waste attention on shuttling messages.” The company appears to be managing AI, but it’s actually desperately protecting the one thing AI didn’t make cheap: whether a person can still care about the right thing. That’s why Fiona concludes: output stopped being the problem; judgment is. What is judgment? Simply the ability to spend attention on the cutting edge. AI plunges everyone into a flood of output. Whoever doesn’t drown and can swim in the right direction wins.

The Most Underestimated Strange Thing: Engineers Started Feeling Lonely

Of the seven items, the one most people skim past but hit me hardest is this one: the team started feeling lonely.

Everyone puts on headphones, grinds away with their own agent, and looks up to find no one around. Anthropic frantically organized “pair programming lunches” and hackathons just to force people back to the same table. Think about what that means: “connection between colleagues” was never a deliberate company benefit — it was a free bonus attached to “we have to collaborate to get work done.” You need me to test a piece of code; I need you to catch a bug. Back and forth, relationships, belonging, craft mentorship all grew naturally within those needs.

Now the agent eliminates the need. I can close the loop alone. That connection disappears with it. For the first time, connection goes from a free byproduct to a “cost center” you must deliberately pay for with time and effort. But here’s a particularly beautiful twist — and a real revelation for all teams: Fiona says when they sat back together, the biggest gain was “watching how other people use Claude Code,” because everyone’s usage is wildly, shockingly different. See that? The reason for gathering has been quietly swapped by AI: people used to gather to “coordinate work” (now agents do that); now they gather to “steal how others think.” The team’s function has shifted from “divided labor” to “cognitive cross-contamination.” We need each other again — not to get the work done, but to learn how to think.

A Dose of Cold Water

Let me throw some cold water so you don’t get too hyped and rush into your office tomorrow to copy “bad vs. sad” and push “JIT planning.”

Remember: these 7 strange things at Anthropic are the effect, not the cause. They are the stress response forced out of a company that slammed headfirst into the wall of “free output.” And you — most likely — haven’t hit that wall yet. Copying their post-crash posture right now is like a healthy person taking fistfuls of medicine for a disease they don’t have. Don’t rush to imitate their rules. First, honestly assess one thing: in your own work, has the bottleneck moved? Where did it go?

That’s the real value of this interview. It doesn’t give you 7 best practices. It gives you a mirror — and one question: When “getting it done” is no longer hard for you, what will you use to prove you’re still valuable?

Fiona already gave the answer: output will become less and less valuable; judgment will become more and more valuable. The machine will fill the world with stuff; the human decides which piece is worth keeping. Those who figure this out early, become early — from a “person who frantically produces” to a “person who decides.”

And the gap between these two kinds of people will grow terrifyingly wide in the coming few years.