X / Twitter profile snapshot for Karan Jagtiani

1/7 People ask me: how do you solely manage cloud infrastructure for a $10M ARR company and still build in public on the side? They assume there's a secret ingredient. There isn't. I optimize for throughput, not optics. 🧵

AI didn't create the problem. It just made faking understanding faster. If someone can't explain the code they shipped, the issue isn't Claude. It's that your review process accepts output without ownership. This is where a lot of teams will get burned. "Works on my machine"

CLI agents feel faster mostly because they hide the work. Less visibility means less scrutiny. And less scrutiny means more slop gets through. This is why "it completed the task" is a useless bar for coding agents. I want to see what changed, why it changed, and where it

The scary part isn't that agents make mistakes. Its that most of them still don't have a real stop button. "Stop" in chat is not control. Its a suggestion. If the model already has tools, memory, and a long enough loop, you're negotiating with your own software. This is why I

Letting an LLM run infra unsupervised isn't "automation". Its delegating consequences to something that can't feel consequences. "Claude wiped prod with Terraform" isn't a freak accident. It's the default failure mode when the tool has write access and you treat chat as intent.

AI slop isn't an argument against AI coding. It's an argument against letting the model drive. Most teams don't have an "AI problem". They have a taste and verification problem. The model will happily generate 200 lines of plausible nonsense if you don't constrain it. The

If your agent costs $3 to run a command you can type in 6 seconds, you're not automating. You're just narrating. The tax isn't the money. It's the habit. You start turning every tiny action into a prompt because it feels "comfortable" than touching the keyboard. Real leverage

Coding getting cheaper doesn't mean software engineering is getting automated. It just shifts the bottleneck to the stuff nobody can autocomplete: system taste, reading a messy codebase, picking the right constraints, and changing it without breaking weird edge cases. Most

People claiming 'AI will replace developers' clearly don't know what they're talking about. Try building and shipping features on a large real-world production project with today's best frontier models. Then get back to me. Most loud voices have only seen todo-app demos.

Just shipped Skyflo v0.6.0. Reasoning is now integrated directly into the agent loop. Every phase of execution can now reason before acting, and that reasoning directly influences mutation decisions. Not “AI thinking out loud.” Actual gated reasoning that feeds into

"Ship fast" isn't a strategy anymore. It's table stakes. The new moat is taste. Knowing what not to build. Saying no to the 5 extra settings. Killing the clever idea that makes the product feel busy. AI made output cheap. Restraint is still expensive. You can spot the teams

AI is quietly turning into an energy business. If Amazon, Google, Meta, Microsoft, xAI, Oracle, and OpenAI are serious about "building their own electricity supply" for data centers, thats not a vibes headline. Its an admission that the binding constraint is no longer just

Everyone sees model curves. Almost nobody models the compute curve. The scary part is not "we are short on GPUs". It is that demand compounds daily while supply arrives in lumpy, slow cycles. Every new agent pattern, every context window bump, every eval framework that runs 10x