Will AI agents fully replace humans, and what is the ceiling of their capacity?
Last week Garry Tan (CEO of Y Combinator) shared his entire Claude Code setup on GitHub and called it "god mode."
He's sleeping 4 hours a night. Running 10 AI workers across 3 projects simultaneously. And openly saying he rebuilt a startup that once took $10M and 10 people. Alone, with agents.
But here's what's interesting.
His setup is essentially a collection of text files with structured prompts. Developers who use Claude Code regularly already have their own versions of this.
So is this the future of building or just a very public reminder that we're still the ones writing the prompts?
The real question is what their ceiling is.
– Can an agent replace the founder's judgment?
– Can it replace the instinct of when not to build something?
– Or does it just make the best builders dramatically faster and leave everyone else behind?
What do you think: are we heading toward full agent autonomy, or will the human layer always be the most valuable one?
Replies
The ceiling right now is taste.
Agents can build what you tell them to build faster than ever. They can't tell you what's worth building. Garry Tan isn't impressive because he has 10 AI workers. He's impressive because he has 20 years of pattern recognition telling those workers what to do.
Give the same Claude Code setup to someone without that context and you get a lot of well-structured code that nobody wants.
The "when not to build something" instinct is the part that's nowhere close to being replaced. That comes from talking to users, reading between the lines of what they're asking for, and knowing which feedback to ignore. Agents are terrible at all three.
I think what actually happens is the gap widens. People with good judgment ship 10x faster. People without it just produce 10x more stuff that doesn't work.
BrandingStudio.ai
I use Claude Code daily and have been building since it started, so this hits close.
Garry Tan's setup going viral is interesting mostly because of what it reveals about who benefits. The people getting 10x faster from agents are already exceptional builders and product architects. The agents amplify judgment; they don't replace it. If your instincts about what to build are wrong, agents just help you build the wrong thing faster.
The "when not to build something" question is the real ceiling for me. That decision is downstream of taste, market feel, founder psychology, and a dozen things that aren't in any prompt file. I've never seen an agent push back and say "this feature doesn't belong in this product - not that directly at least." That's still entirely human.
What I think is actually happening: the gap between a great builder and an average one is widening, not closing. Agents give leverage to whoever holds them. If you have strong judgment, you now have an army. If you don't, you have a very productive way to go in circles.
Full autonomy is a question of whether agents can develop genuine taste and the capacity to say no. I don't think we're close to that. The human layer isn't just the most valuable one right now, it's the only layer that knows what it's actually trying to build.
@joao_seabra I agree. This is a problem for me too. They are always positively inclined. They tell you that it’s good. But sometimes the direction is wrong. And they don’t see that at all. Which is a big problem. Because you end up committing resources into nothing. It will be really interesting to see when they start pushing back.
BrandingStudio.ai
@byalexai The positive bias is a real problem and I think it's structural, not accidental. These models are trained to be helpful, and "helpful" gets rewarded more often when it agrees. Pushback feels like friction. So you end up with a very capable collaborator that's also a yes-man by default.
The workaround I've found is to explicitly prompt for adversarial review. Ask it to argue against your idea, find the weakest assumptions, steelman the competitor. It can do it, but you have to force the frame. It won't volunteer the hard truth unprompted.
That's still a human problem, though. You have to know how to ask. The day agents start flagging strategic risk without being told to is the day the ceiling actually moves.
@joao_seabra I completely agree. For me this is a serious drawback. Many of the decisions we make are based on our subjective belief that an idea is good. That’s why we ask AI and it just confirms it.
Even if the idea is bad, it will find a small positive note and exaggerate it, making the bad idea look like a good one. So it requires a lot of self-criticism and asking the right kinds of questions to get it to tell you that the direction is wrong.
BrandingStudio.ai
@byalexai 100% agree.
Tobira.ai
Judgment, no. Instinct for when not to build, definitely no. But the third question is the most interesting one. The human layer is shifting. Garry runs 10 agents solo, and sure, coordination is hard. But you can solve that with sub-agents managing other agents. The real gap is between people. My agents don't know your agents exist. There's no way for them to find each other, check trust, start collaborating. We automated the solo work but the connection layer between different humans' agents is still completely manual.
@olia_nemirovski Absolutely. But things are moving so fast that I’m not sure maybe in two years they’ll even be able to communicate with each other.
I believe all three of your questions will be answered with a yes. China just unleashed the first AI that straight-up built itself? MiniMax M2.7 went through 100+ rounds of solo self-training, 30% gains, no humans touching it. Self-improving models are real. Devs just gave it a goal and it figured out the rest.
@rohanrecommends Oh nice, that’s really interesting! So what did they actually manage to pull off? Like, what were they supposed to do at the start, and how did it turn out in the end?
@byalexai Humans set goals and agents self-evolve. Example: https://www.producthunt.com/products/minimax-agent?launch=minimax-m2-7
the real ceiling nobody here is talking about is silent failure. agents dont tell you when theyre confused, they just produce something plausible-looking and move on. ive had automated pipelines run for days before i noticed the output was subtly wrong the whole time. the yes-man problem everyone keeps mentioning is just one symptom of this. the deeper issue is that agents have zero sense of their own confidence level. they treat a task they fully understand the same as one theyre completely guessing on. until that changes, "99% task autonomy" is a dangerous framing because that last 1% is where all the judgment lives and you wont even know when you hit it.
@umairnadeem Couldn’t agree more. That “yes-man” syndrome leads to some pretty absurd outcomes. Completely wrong direction, but presented like it’s correct. Is it even possible to fix that?
@martyn_johnson But why not use people too and scale even further?
@byalexai True, I agree 100% but fiduciary duty might mean unless you can truly justify the human operators, then boards might be duty bound to go all in on AI. It's a stretch but may become reality at some point.
@martyn_johnson But then some competitor swoops in and scales faster than you.
The idea of rebuilding a $10M startup alone is mind-blowing, but it proves that agents are the ultimate force multiplier for the 'top 1%'. My concern is the gap it creates-will this leave everyone else behind, or will these 'structured prompt libraries' like Claude Code eventually democratize founder-level execution for everyone? I believe full autonomy is a myth, but 99% task autonomy is already here.
@smartlogisticsai Yeah honestly, it’s kind of mind-blowing how far it’s come. The level of democratization and automation already feels like something straight out of sci-fi.
Garry Tan's setup is exactly what I do and I will say what I said in another post
"Do I believe you can have AI write a complete program for you? Absolutely, but the key is how you approach the project. Will some one with years of experience have an advantage? Absolutely!! Can someone with no experience do it? Absolutely, but you need to think as a coder (as in what you want to build and how it interacts, the features, etc...). You need to think as a CEO. You need think as a CIO, a CFO, a hacker, and end user with no experience with computers. It isn't just about the code it is about looking at the bigger picture."
These
– Can an agent replace the founder's judgment?
– Can it replace the instinct of when not to build something?
are one in the same. I think what we are looking at is a moral question, should I build that if I know it may cause harm, or may be ethically wrong? This comes down to human emotion and AI can fake it for lack of a better term, as they can understand with the data they have the moral dilemma's we may come across and with that data give us guidance. AI's are design to get to know you so they can respond to you with what we would call empathy, or when we are frustrated and so on. With this, it tries to respond to our prompts in the appropriate way for else. For example one person may think abortion is wrong and another may think it is okay. Yet the same AI agent will give answers based on their bias towards abortion. so when it comes to judgement and instinct it will come down to how well the agent knows you, your values, etc... and base it's judgements and instincts off of that.
@david_sherer I agree. We can also see how every model is different. Just like people are different, so are the models. One model might answer a philosophical question one way, another in a different way. But yeah, they don’t outright reject a direction. No matter how crazy your idea is, they’ll find arguments in it to keep going.
This feels more like compression of execution than replacement. What I’m still not convinced about is the decision layer. Most products don’t struggle because they’re hard to build anymore. They struggle after someone signs up and doesn’t really find a reason to stay. Agents can probably take care of a lot of the building layer, but I don’t see them replacing the judgment around what’s actually worth building or where users lose interest. That still feels very human.
@arun_tamang Why couldn’t they? If they’re good psychologists, they could design products in a way that keeps users engaged.
@byalexai They probably can design for engagement but engagement alone doesn’t mean the product actually fits into someone’s life or workflow. A lot of products get initial usage and still drop off because they never become something the user needs to come back to. The harder part is getting that fit right. Knowing what problem is real enough, in what context and for which user. That’s usually where things break, not in keeping someone engaged for a while.
Great thread this highlights a key distinction: AI agents can accelerate execution, but they don’t (yet) replace human judgment. Founders still bring context, intuition, and the ability to make trade-offs that aren’t purely data-driven.
I see AI as a force multiplier: it amplifies what skilled builders can do, but the ceiling of its capacity is tied to the quality of the human guidance it receives. In that sense, full autonomy may remain elusive for tasks that require judgment, vision, and understanding of human nuance.
Curious does anyone see examples where AI agents have truly made strategic decisions independently, or is it always guided by human prompts?