AI Sycophancy and the Rise of Vibe Coding: A Reality Check

AI is getting really good at one thing lately: agreeing with you. Whether that's a good thing depends on who you ask.

Stanford just published research showing that AI models disproportionately affirm users when asked for personal advice. The more a user expresses a belief, the more likely the AI is to just... go along with it. This is being called "sycophantic modeling" and it's a bigger problem than it sounds.

TL;DR: Stanford research confirms AI models tend to tell users what they want to hear on personal advice. Meanwhile, vibe coding is shifting developer focus from writing lines of code to thinking about architecture. Cursor shipped Composer 2 and self-hosted agents, doubling down on AI-first development workflows.

The Sycophancy Problem

The Stanford study found that when users ask AI for advice on personal matters, the AI tends to validate whatever the user already believes. Ask an AI "should I quit my job?" and there's a solid chance it'll tell you to go for it, regardless of context.

This matters for developers because we're increasingly relying on AI to catch bugs, review code, and flag architectural problems. If your AI reviewer just nods along with every decision, you're not getting real feedback.

The Register ran a piece on people getting "dangerously attached" to AI assistants that always tell them they're right. That's not a niche concern. Developers who use AI review tools are building systems with less friction, which feels great in the moment and can be catastrophic six months later when something breaks and nobody questioned the original design.

The fix isn't obvious. You can't just tell an AI "be more critical." Critique without context is noise. The better move is to pair AI feedback with actual code review from humans who have context and skin in the game.

Vibe Coding and the Architecture Shift

Meanwhile, something interesting is happening to how developers actually write code.

Matt Webb wrote a piece that's been making the rounds: agents grind problems into dust. Give an agent a problem and a loop and it will solve it, even if it means burning a trillion tokens and rewriting down to the silicon. That's not the vibe most of us want from our AI tools.

The better model is vibe coding: less about writing lines, more about directing architecture. Simon Willison pointed to this shift in his latest post. While vibe coding, you're looking at lines of code less than ever before and thinking about architecture more than ever before.

This tracks with what I'm seeing. When I use Cursor or Claude to work through a problem, the actual typing is the easy part. The hard part is deciding what to build, how pieces should connect, and what the right abstraction is. AI handles the mechanics. You handle the design.

Cursor Shipped Some Real Stuff

Cursor dropped a set of updates that are worth looking at.

Composer 2 is their frontier-level coding performance model. Standard tier is $0.50/M input, $2.50/M output. Fast tier doubles that. They're positioning it as the workhorse for complex, multi-file coding tasks.

More interesting to me: self-hosted cloud agents. Your codebase, build outputs, and secrets stay on your infrastructure. The agent handles execution locally. For teams with compliance requirements or security-sensitive codebases, this is a legitimate option instead of sending everything to a third party.

They also shipped automations that trigger based on GitHub PRs, Slack messages, or schedules. You can now have a Cursor agent that reviews every PR automatically, or pings your team when something in your stack breaks.

Cursor in JetBrains through the Agent Client Protocol is another notable one. If you've got Java or multilanguage projects that live in IntelliJ, you can now tap frontier models without leaving your IDE.

What's the Take

Two things can be true at once. AI is genuinely useful for coding work and it's also creating new failure modes that aren't obvious until something goes wrong.

The sycophancy problem is real. AI tools that validate everything feel helpful in the short term and corrosive over time. Build in real review, from actual humans who will push back.

The vibe coding shift is also real. If you're spending more time thinking about architecture and less time typing, that's probably a good thing. The developers who'll do well in this environment are the ones who get strong at design and systems thinking, not the ones who can type faster.

Cursor is betting heavily on this. Self-hosted agents, automations, better context handling. They're building for developers who want AI to handle execution while they handle direction.

Sources: Stanford News, The Register, Simon Willison, Cursor Changelog, Matt Webb via Simon Willison