Alibaba just tested AI coding agents on 100 real codebases, tracking them over 233 days each. The agents failed spectacularly. And if you're holding AI stocks priced for perfection, you might want to pay attention.
The benchmark is called SWE-CI, and it's the first test that measures long-term code maintenance instead of one-shot bug fixes. Each task tracks 71 consecutive commits of real code evolution. The kind of stuff human developers actually do: maintaining a codebase over months, handling dependencies, not breaking things as requirements change.
Turns out AI is terrible at it. Passing tests once is easy for these models. Maintaining code for eight months without breaking everything is where they collapse.
Why does this matter for investors? Because the entire AI trade for the past year has been predicated on the idea that AI will replace huge chunks of human labor, starting with knowledge work like coding. If AI can't handle long-term software maintenance—one of the most automatable knowledge tasks—what does that say about AI's near-term economic impact?
Let's be clear: this doesn't mean AI is useless. AI coding assistants like GitHub Copilot are legitimately helpful for writing boilerplate code, suggesting functions, and speeding up development. But there's a massive difference between "helpful assistant" and "replaces the developer."
The market has been pricing AI stocks like they're going to replace developers. The SWE-CI benchmark suggests that timeline is a lot further out than people think.
Here's the practical impact. If you're holding NVIDIA, Microsoft, Google, or any of the big AI infrastructure plays, this isn't necessarily a sell signal. Those companies are building the rails. Even if AI adoption is slower than expected, they're still going to benefit.
But if you're holding speculative AI software companies that are valued based on replacing human labor in the next 2-3 years, this is a yellow flag. The technology isn't there yet. It might get there eventually, but "eventually" is not what's priced into these stocks.
The AI skeptics have been saying this for months: AI is great at demos, terrible at production. One-shot tests make AI look amazing. Long-term reliability is where it falls apart. SWE-CI is just the latest proof.
Does this mean you should sell all your AI stocks? No. But it does mean you should temper your expectations. The AI revolution is real. The timeline and scope might be overhyped.
AI stocks have already been under pressure for months. Tech had a rough 2025. This is just one more data point suggesting that the AI trade got ahead of itself.
If you're a long-term investor, none of this changes the thesis. AI will eventually transform knowledge work. But if you're trading on near-term hype, this is the kind of news that can accelerate a selloff.
The people who bought AI stocks in 2023 are still sitting on massive gains. The people who bought in 2025 at the peak are underwater. That's the risk of chasing momentum in a sector priced for perfection.
AI limitations are real. Plan accordingly.
