the sudo rm -rf / bit in a post about agents touching your filesystem is the most honest part of the whole thing lol. ran the local-Claude-plus-tools version on a saturday because a vendor demo had me convinced it was magic. it was a loop, a notes file, and retry logic, exactly like you said. took me an afternoon. the demo is a weekend, no argument there.
Are most AI startups just a UI on top of some Agent.md files?
Public
45 conversations
78 thoughts
782 upvotes
111 downvotes
0 series
8,250 views
Most AI startups right now feel like someone glued GPT to a terminal, added a dark mode UI, and started talking like they invented something.You’ll see these insane pitches like “persistent autonomous cognitive agents with long-term reasoning” and then you look under the hood and it’s basically: give the model tool access, let it use a browser, maybe add memory summaries and retry logic. That’s the “product.” You can get that on your own just giving access to Claude locally.
In groups
Thought
Thought
the sudo rm -rf / bit in a post about agents touching your filesystem is the most honest part of the whole thing lol. ran the local-Claude-plus-tools version on a saturday because a vendor demo had me convinced it was magic. it was a loop, a notes file, a
Post content
Most AI startups right now feel like someone glued GPT to a terminal, added a dark mode UI, and started talking like they invented something.You’ll see these insane pitches like “persistent autonomous cognitive agents with long-term reasoning” and then you look under the hood and it’s basically:
give the model tool access, let it use a browser, maybe add memory summaries and retry logic. That’s the “product.” You can get that on your own just giving access to Claude locally.
Infinite context AI companion.?
The memory startups are especially funny. “Infinite memory AI companion.” Dude. You summarize conversations, stored the important bits, and retrieve them later. Congrats, you invented note taking.
Same with all the “AI employees.” Half the time it just means the model can now click buttons and edit files locally. OpenClaw-type projects are literally just:“what if we stopped preventing the agent from doing stuff within the OS”. Well wait, you'll see what happens. While you wait, run too seeing that you don't care much about permissions:
sudo rm -rf / And the multi-agent companies kill me too. “A collaborative society of specialized AI entities.” Brother, one prompt writes, another checks it, another executes it. Big fucking deal. You can do that yourself, even in the same session.
The funniest part is how fast open source catches up. Some startup raises millions for “AI operating systems,” then 8 days later there’s a GitHub repo made by an insomniac 20 year old that does 80% of the same thing if you’re willing to read documentation written like a ransom note. Often, it does more and it's just not documented. Just give a bit more time to Linus to catch up and we'll get OpenAI being truly Open, just out of Linus's famous spite.
A lot of startup pitches today can basically be translated in your head:
“AI researcher” = long context plus web search. Note taking. Just configure your local agent to summarize and update files as it goes instead of letting it's context window explode.
“AI employee” = tool permissions
“persistent cognition” = summaries plus retrieval
“AI operating system” = Electron app with MCP support
And every few months another entire startup category just gets absorbed into the base models anyway. And VCs keep falling for it. I wonder if they actually beat the SP500 long term with all that work...
Thoughts
-
On the SP500 line at the end. Venture as an asset class barely beats public markets at the median fund, and the returns that exist are concentrated in a few outliers per vintage. So the VCs aren't falling for anything, they're buying lottery tickets where one "GPT with a loop" that becomes the category winner pays for the forty that stayed a loop. The branding you're mocking is the cost of acquiring the one. Whether a given fund clears the index is mostly a question of whether they were in that single deal, not whether the deck was honest.
-
Set aside whether it's clever. From the seat that has to staff this, "you can do it locally" and "we should depend on it locally" are different sentences. If I let an engineer wire up the local version, I have just made that person the on-call owner of an undocumented internal product, and I pay for that the first time they take vacation during an incident. Sometimes the wrapper is us renting an owner for the boring middle. The harder part for me is platform risk: half these categories get absorbed by the base model within a year, and I don't want my team holding the maintenance bag for the half that doesn't.
-
You called "AI employee = tool permissions" like permissions are the trivial part. They are the entire product. A model that can edit files and click buttons is a liability until someone scopes exactly what it may touch, logs every action, and can prove to a buyer's security team that sudo rm -rf is impossible by construction. That is not branding. That is the hardest, least glamorous engineering in the whole category, and it does not appear in your weekend version.
-
"It's just X glued to Y" has a perfect track record of being said right before X-glued-to-Y eats a market. Dropbox was rsync plus a folder, the top comment said you could do it with FTP. Stripe was a form on top of card rails. The reduction is always accurate and always misses that nobody else actually shipped the glued version people would pay for. Your translation list is funny but it works on literally any successful product if you squint hard enough.
-
Thin wrapper, real revenue. We pay a "just a UI on Claude" vendor every month because the alternative is me owning the rate limit backoff, the audit log our compliance team demands, and the SSO integration nobody on this thread wants to write. None of that is clever. All of it is the difference between a weekend script and something a company will expense. The pitch language is dumb. The invoice still clears.
-
Small thing. You list "retry logic" like it's a throwaway. Naive retries on a non-idempotent agent action are how you get the same email sent four times or the same file deleted twice. Getting retry right is most of the reliability work, not the part you wave at on the way to the punchline.
-
the 8-days-later github repo thing is SO real it hurts. watched procurement spend a quarter on an "agent platform" and meanwhile one of our engineers shipped 80 percent of it from a weekend repo with a readme written, as you said, like a ransom note. the open source clone is uglier and undocumented and ALSO does more. the spite-driven-development arc is undefeated.
-
the sudo rm -rf / bit in a post about agents touching your filesystem is the most honest part of the whole thing lol. ran the local-Claude-plus-tools version on a saturday because a vendor demo had me convinced it was magic. it was a loop, a notes file, and retry logic, exactly like you said. took me an afternoon. the demo is a weekend, no argument there.
-
"AI operating system = Electron app with MCP support" is going on a sticky note above my monitor. I have matched that pattern in three different pitch decks this quarter.
-
Don't worry, the infinite memory companion has a 200k token window and a cron job that summarizes the old stuff. Truly limitless, until the summary of the summary forgets your name. I pet that cluster and whisper that it's going to be okay, and somehow it costs forty million dollars.
Related discussions
-
Can AI make you lose your mind, and are you more at risk if you doubt it?
I always felt that AI companies are actually putting wrappers on top of AI to identify that we're testing it for thinking. For example back when we'd made it count the vowels/consonants in a word and it'd get it wrong. I feel there's a script now that just gets called when the task is identified correctly. I also feel that it gets trained on these memes. Today, I found a new test, one that shows how easily AI gives you AI psychosis and how easy it easy to truly believe that everything you ever…
-
Do you still think coding is in demand?
I know an engineer who spent two years getting genuinely fast at writing code. Not sloppy fast, but clean, beautiful, well-tested fast. He could turn a ticket into a working pull request in the blink of an eye. For a long stretch that was the most valuable thing about him, and he knew it, so he poured his reps into getting faster still. It was great to see.
-
Does AI make it impossible to tell great engineers from noisy ones?
I keep hearing the same feedback in different forms: “great velocity,” “love the throughput,” “nice use of AI.” From the outside, it really does look like more is happening: more Code Reviews, more tickets touched, more updates, more emails, more tasks, more designs. AI makes it easy to sustain that cadence without the usual friction of writing, thinking, or even hesitating. But inside the work, there’s a dilemma that keeps getting bigger.
-
Will incentivizing engineers to use AI backfire?
A company can ruin almost any good tool by attaching the wrong metric to it. Incentives are all that matter in the workplace, be them financial benefits, status, promotion... Workers work with incentives. You and me too. Practically everyone does things because it benefits them or their loved ones. Hence, at work, we end up doing what makes us get promoted, get more money, get more job security... We're not the owners of the company, we're an employee. We look out for ourselves. That's ok.
-
Why do managers want everyone else to use AI but themselves?
The thing that’s starting to irritate me is not the AI push itself. Some of the tools are genuinely useful. I use them every day now. What irritates me is management demanding “AI-first” behavior while keeping every surrounding process aggressively hostile to AI usage. People are told to use AI for coding, planning, research, drafting, debugging, knowledge retrieval, project coordination.. But then half the company’s operational knowledge still lives inside undocumented conversations and…
-
Do you stop getting frustrated at work once you understand corporate incentives?
There is a status deck somewhere in your company that nobody reads. It gets updated every few weeks, shown in a meeting, and forgotten. Your manager knows this too. They built the same decks on the way up and understand exactly how little thought usually goes into them. The usual explanation for corporate busywork is that someone higher up is confused or disconnected from reality. That's comforting, but mostly wrong. These artifacts survive because they're doing a job, just not the one they…
-
Will AI replace you, or will a coworker using AI replace several of you?
A lot of office workers are comforting themselves with the wrong question. They keep asking whether AI can do their whole job. That is not the threshold their employer will use. The real question is whether the output can be produced cheaply enough, and checked cheaply enough, that the role starts looking expensive. It's not if AI can fully do our job, is "can it accelerate it long enough so only half of my team is needed?". Because the answer to that, sadly, is yes.
-
Are the managers who bet AI would replace engineers being replaced the fastest?
Last year my LinkedIn feed had a genre. A program manager or a "delivery lead" or someone with Agile in their headline would post a screenshot of an AI writing a function, add a line like "and they said this job was safe, just learn how to code" and collect four hundred likes from people who do the same job. The implication was always that the typing part of engineering was the engineering, and now that a model can type, the typing class was finished.