Claude Blue Is Spreading Across Silicon Valley

Conversations with a senior AI engineer at Meta and a Silicon Valley founder

Claude Blue is a phrase I made up, but it describes something real: a wave of AI-induced depression spreading through Silicon Valley.

AI hit the software industry in 2025. Back then there were two camps. One said coding should be handed over to AI entirely. The other said a human should at least review everything. Now, in March 2026, that debate barely matters. AI already writes more code than humans can review. We walked straight into the transition.

I expected the same wave to reach office workers in 2026. Coding had shown the way, so I assumed jobs built on documents and data analysis would be next. But I asked about fifteen people I know in the US and Korea, in all kinds of jobs, and none of them felt that AI was slowing them down or causing problems yet. They all said it was still something people could control.

We argued about this at my company too. Everyone agreed that AI would soon change office work at a basic level. We disagreed on when. I thought it was still early. I'm a little embarrassed to admit it, but I had never personally been held up at work because of AI, and neither had anyone around me. One colleague disagreed. He said office work had already changed dramatically and the damage had started: people were losing track of the reasoning behind documents and decisions that AI had produced. He didn't blame anyone for it. To him it was simply a cost of the transition we're in.

Then today I had a video call with an older friend from college who now works as a senior AI engineer at Meta, and he convinced me. At the leading edge of Silicon Valley this was already happening, and hearing his concrete examples made it real for me. Within weeks, or a few months at most, office workers will be pushed to become AI-native too. Nobody will write documents by hand or click through websites taking notes. Working in a terminal and directing agents, with Claude Code or Cowork, will be normal. And office work will take the same transitional hit that coding did.

What struck me even more was that he, someone at the very front of AI engineering at Meta, was having a hard reality check of his own. He told me he now feels a strong need to develop a founder's instincts and build a business. He had always been interested in startups but never pictured starting one. These days, he said, even his side projects are driven less by fun and more by anxiety. He feels the day is coming when he'll have to run his own company.

The same day, I met a friend my age who runs a startup in Silicon Valley for coffee, and to my surprise he sounded almost the same. An engineer at a Big Tech company and a startup founder are in very different positions. Both were feeling the same existential crisis about AI, and the same desperate sense that there's no stopping it.

Here is what we talked about, topic by topic.

Editor's note (April 2026): Thanks to so many readers, this essay went viral and led to an official community event with Anthropic's Claude Community: Claude Bloom ("From Claude Blue to Bloom"), Seoul, April 14, 2026.

1. An AI-native day at Meta: working only through agents

He uses Claude Code at work and OpenAI's Codex at home. The reason is simple. At work his only choices are Claude Code and Gemini, and Claude is far better at coding. At home, for personal projects, he needs image generation, which Claude Code doesn't offer. And the $200 Codex Pro plan gives him what amounts to unlimited tokens. Claude Code Max also costs $200, but heavy use costs extra. At work he goes through about $2,000 a month in Claude Code tokens, which he couldn't pay out of his own pocket.

His setup was fascinating. He built his own VS Code extension that automatically grows a knowledge graph in Obsidian, the note-taking app. Every conversation he has with AI becomes an Obsidian note, links itself to his existing notes, and the graph keeps growing. Obsidian used to mean writing and linking notes by hand. Now the conversations with AI are the raw material, so it's much easier to get started. Meta has about 70,000 employees, and more than 10,000 of them have joined the internal Obsidian users' group.

What surprised me more was that every productivity tool inside Meta is hooked up to agents. Google Drive, Docs, Sheets, email, Jira, Confluence, the wiki, whiteboards: all of it is connected. In his words: "I only work with agents, and the results pile up in each tool in whatever shape I want." He never has to open the spreadsheet or Jira. He stays in the terminal and the databases fill themselves.

One example showed how far along this is. In January, for his first year-end review, he asked the internal AI review tool to "summarize everything I've done since I joined." It went through every report and document he had written and produced a huge list. All he had to add was his own reflection, and his self-review was done.

2. "The infrastructure is wrecked": the AI code crisis up close

AI hasn't only made life easier at Meta. The way he works shows both sides.

When he gets a new task, he opens two agents at once. The first does the work. He gives it the goal and tells it, "Go, get started." He doesn't fully understand the task yet himself. He just hands over the objective and lets it run. The second is a learning agent that explains the task to him. After the first agent has run for a day or two, there are close to ten thousand lines of code. Meanwhile he learns the task from the learning agent, then has yet another agent analyze the code. The problems he finds are mostly small fixes, and the agents make those too.

The trouble is that people can't keep up with that speed. He put it bluntly: "Meta's infrastructure is completely wrecked right now." Too many people push code they haven't read. AI-generated code gets committed without enough review, and the stability of the infrastructure is suffering badly. This is the transitional damage of the vibe-coding era.

He thinks the same thing will happen outside engineering. Say a company that isn't in software has 100 people. First, someone will start a company where three people do the same work with agents, charge less, and put the old company out of business. Then those three-person companies will hit their own crisis, because they can't handle everything the agents produce. The ones that can't cope will fail too. Only the companies whose three people can really orchestrate agents will survive.

3. From vibe coding to agent orchestration

He says people in Silicon Valley hardly say "vibe coding" anymore. The term now is "agent orchestration." Vibe coding means tossing out a loose request like "make me a game." The approach today is much more structured. For a game, you write a very detailed spec: the genre, how far a character can level up, how the enemies get stronger. Then you take the things you keep asking for and package them into reusable pieces such as skills or tools. How well you feed those to your agents decides how good the result is.

Being good at your job now means being able to get AI to produce the result you want. As I understood him, the human's role is shrinking to the beginning and the end: sensing what should be made, and judging whether the result is any good. Only people can do those two things. Beyond that, the people who do well will be the ones who can say exactly what is wrong with a result and how to fix it. At the company level it gets much more complicated. You break big projects into small pieces and decide which parts AI can check on its own, which parts a person has to look at, and which parts AI can move forward without asking. Agents cost tokens too, so getting the most out of both people and agents becomes a core skill in its own right.

4. Non-engineers are in it too

At Meta, he said, AI is used well beyond engineering. Program managers, HR, sales: practically every role runs on AI now.

At his child's birthday party, he met an Indian dad who manages sales data at Apple. This dad had moved his whole team to working through agents. It's a data science team that turns data into insights for management decisions, and they handed all of that to agents. People got time back, and they used it to take on more work as a team.

On Meta's internal Facebook, people in HR and People teams who aren't engineers regularly post things like, "I'm not an engineer, but I built this. I think it's good, give it a try." One HR team built an agent that pulls data from every software tool engineers log into, so it can see what employees are working on and score them with its own system.

The same goes for people who spend a lot of time in meetings, like salespeople. You put the key messages in a spreadsheet, hand it over, and the slides get built for you. He doesn't make slides by hand anymore either. He drops in the paragraphs and charts he wants and says "make it." After a few rounds of "make this table a bit bigger," he has something he likes.

His conclusion: "I don't think there's anyone at Meta who isn't using AI right now."

5. "Wow, this is going to replace me": an engineer at the frontier faces facts

He has a PhD in physics and engineering, and he spent years building his expertise, first at one of Korea's biggest conglomerates and then at Meta. When someone like that has this kind of reckoning, it carries weight.

"I had my own methods, like how I structured writing and how I did visualizations. That was my edge. But when I ask AI to organize something, it writes so well. Out comes a polished piece." Know-how he had built over years, even decades, was his advantage, and AI replaced it in an instant.

He could even say when it happened. Until early February he was excited. Every new AI feature was "wow, amazing, so fun." Then Claude Opus 4.6 and Codex 5.4 came out, and his mood changed completely. "That 0.1 difference was the point where I started to feel threatened." After using the mid-January release seriously from February on, he thought, "Wait, what? I don't have much time left."

He was honest about how he felt: denial, frustration, anger, and only in the last few days a slow turn toward acceptance. The whole cycle is surprisingly short, which tells you how fast the technology is moving. "It used to be that you had to really use a new technology before you could judge it. Now you watch two hours of YouTube, try it once, and go, wow, this is going to replace me, I'm done. It happens in a day."

He said what he's proudest of is his ability to learn. He's confident nobody can beat him at that. But against the volume of what the agents turn out, he is losing.

6. The base-model war is over, and Big Tech wins the ground

The conversation moved to how the AI industry is structured. His view was clear: the fight over base models is winding down, and the outcome is already settled.

Only four or five companies have base models good enough to actually replace people's work: Anthropic (Claude), OpenAI (GPT), Google (Gemini), xAI (Grok), and, if you count it, Meta. People might play around with Chinese models like DeepSeek or Kimi, but only those four or five can do real work at a level that replaces people.

The key point is that only these companies get to build the engineering layer for agent orchestration. Because they own the models, they can build the systems that let agents talk to each other and create value in each field without worrying about margins. A company that builds on top of someone else's model, like Perplexity, has a ceiling. If Google pushes AI a little further into search, Perplexity is in trouble right away.

Meta's strategy is different from the others. Meta builds AI for its users. It has to be strong but cheap enough to run while Instagram users scroll through high-resolution video. That's a very different product from business productivity tools like Claude or GPT. So inside Meta nobody hesitates to use outside AI tools. As he put it, "The AI we make is for our customers to use, not for us."

7. Organizations that don't change get eaten

He used to work at the headquarters of a major Korean conglomerate. Same AI engineering job, completely different life, he said. The big difference was how many layers of permission you need to get anything done. At the conglomerate, you get your manager's approval, your manager gets their manager's approval, then there's a contract, then preparation, then a report on the preparation. At Meta you just build it. You post it, tag the relevant people, and ask, "I did this, OK?" If one person signs off, it ships.

Part of that is Korean versus American culture. The deeper difference is manufacturing conglomerate versus software startup. You have Korean conglomerate culture plus a heavy focus on security on one side, and Meta's software culture plus startup culture plus American culture on the other. That mix makes life completely different for people doing similar work.

He was frank about organizations that don't use AI. "If they're not using it now, there's a reason, and they'll keep not using it." On the surface the reason is security, but underneath there are vested interests. An affiliate company builds and supplies the in-house AI tool, that arrangement is already in place, and bringing in Claude Code from outside would upset it. The people at the top won't allow that.

His analogy stuck with me. He pointed to NCSoft, the Korean game company behind the long-running online game "Lineage". NCSoft hired AI talent at scale and invested heavily, but its business model, selling expensive in-game items, never changed. The AI features got tacked on in places nobody uses. The formula that got them this far is exactly why they won't change when something new comes along.

"Eventually, companies like yours will eat companies like that."

He also looked back on his move to Meta. "I timed it really well. With the times changing like this, if I'd stayed, I still wouldn't be able to use AI." At the conglomerate he had worked by copying and pasting code back and forth with an internal chatbot. Compared with working directly with a coding agent, he puts the productivity gap at about five times. "What took me three months there, I can do in a week now."

8. The next generation's only weapon: an entrepreneurial mindset

Then we talked about raising kids. He has a young child. He thinks a huge transformation will play out and settle over the next five years. By the time his child is in middle school, the world will already have been reorganized, and there will be clear examples to plan a future around. So in his view, his child was actually born at a good time.

There's a catch, though. "The next generation has to be entrepreneurial. Kids who don't think like owners won't find a place." So he works hard not to crush his child's spirit. When you scold a child, their spirit shrinks. A child whose spirit has shrunk easily turns into the kind of kid who only learns to ace standardized tests, and that kind of kid may have no answer in the world that's coming. Helping his child build confidence and an entrepreneurial streak is the one rule he follows as a parent.

At the same time, he expects the next five years to be rough as the world changes. People in jobs with a high chance of being replaced by AI need to do something, anything. "This year really matters. It matters so, so much."

9. The two faces of Silicon Valley: depression and big dreams

That same day I had coffee with a friend my age who is building a startup in Silicon Valley. He has a Stanford MBA, founded his company in San Francisco in 2024, and has built and shut down more than forty products since. He was in Korea briefly for work, and what he told me was remarkably close to what my friend at Meta had said.

"Silicon Valley is actually really depressed right now. It makes me want to leave." He said people there even have a name for it: "Claude depression." People feel, painfully, that AI will eventually take every human job and role. Meta's recent layoff of 16,000 engineers and its freeze on all new-grad hiring aren't abstract headlines to them. They're happening right next door.

His company's mission was "give everyone the power to build software." The idea was to close the gap between developers and everyone else by giving non-developers the ability to build with AI. In 2024 that was the right bet, so apart from himself he hired only software engineers. Now he worries that he went all in on one thing and that the company is missing other skills. He said the team talks about it openly.

The line that hit me hardest: "You sell customers on today and investors on ten years from now. But we all know that in ten years we'll all be lying in bed." Everyone around him feels this way, and a strange gloom hangs over the whole Valley.

He still doesn't want to leave, though. When he first went abroad for his MBA, he found the sameness of Korean society stifling and wanted diversity. Looking back, Silicon Valley is diverse in terms of race, but in the end everyone is a startup or AI person, so it's uniform in its own way. Still, in Korea it's hard to dream big, and Silicon Valley is full of crazy people with big dreams. He wants to keep living in America for that inspiration.

What I found interesting was that this gloom pushes people's imaginations toward things far from AI software. If he were investing, he said, he'd put money into Korean food. Netflix in the US always has Korean shows on, and as Americans get more into K-culture, they naturally get curious about the food they see on screen. The problem is that Korean food in America doesn't feel young. If you brought over the feel of Seoul's trendy neighborhoods like Gangnam or Seongsu as is, it could compete. The American market mostly has old chains like Chipotle and McDonald's, and there's real demand for someone new. Someone at the frontier of AI software seeing a thirty-year opportunity in the industry furthest from tech seemed to me to sum up the strange mood in Silicon Valley right now.

There was a darker story too. AI will wipe out a lot of jobs, and historically, when countries have needed to create jobs fast, the biggest driver has been war. So in some corners of Silicon Valley people say, half joking and half serious, that the smart trade is to go long on defense companies and short AI startups. The romantic version is that people who lose their jobs to AI pour into VR and entertainment. The tragic version is that countries go to war and people die. Several wars have in fact broken out sooner than anyone expected, so it doesn't sound like just a joke anymore.

A senior AI engineer at Meta and a startup founder in Silicon Valley. Big Tech and a startup are very different places to stand, but what the two of them felt was remarkably alike: dread about the future AI will bring, the desperate sense that it can't be stopped, and the belief that an entrepreneurial mindset is the only answer. That was the mood in Silicon Valley in the spring of 2026.

10. Relationships are still something AI can't replace

In the age of AI, everyone wants to build a global business. That raises an obvious question: do you still have to be there in person? I asked both of them.

My friend at Meta was clear. "Business still runs on relationships." Just the day before, a VC had asked him for advice over Zoom. He thought he had talked fairly openly, but after talking with me in person today, he felt how different the depth was. "Building relationships really matters. But can you build a relationship over Zoom? Don't you have to share a meal?" Seeing each other's faces and eating together puts people at ease, and he doesn't think AI changes that.

My founder friend said the same thing. He calls himself an introvert who's bad at sales, yet he found his first customers through the Stanford alumni network. Stanford people are always curious about startups, and many signed up simply because they wanted to learn too. And in the middle of Silicon Valley's AI depression, he came to Korea to see old friends and spend time with family. However global software gets, and however remote work becomes, some things only get filled by meeting people.

When I think about it, being able to call up a friend at Meta and have a conversation this deep is itself an asset nobody can see. Other people would have to pay just to interview someone at a Big Tech company in the US. Relationships like this can have a huge impact, whether or not they ever turn into business.

After a day of talking with a Big Tech engineer and a Silicon Valley founder, one thing became clear to me. At the start of this essay I wrote that none of the fifteen people I asked felt AI holding them back at work, and I wondered whether it was just too early. Today's conversations changed my mind. If we don't feel held back, it's probably because we aren't using AI enough yet. People who work in Silicon Valley are already feeling Claude Blue. I don't see why we would be any different.

If anything, maybe we should push our AI agents as hard as they do and keep looking for the bottleneck until it wears us out. We're in an age where, if you don't have at least three agents running while you're awake, you should feel your life slipping away. I felt an urgent need to get at least one more running today.

I'm not an AI expert, but I'm always up for a coffee chat. If you'd like to share your story or hear more about other people's experiences, reach out on LinkedIn.

Originally published in Korean on Brunch, March 18, 2026.