Why we stopped using SDKs
X
A dev shop says AI coding tools made SDKs pointless—wrapping an API is now as much work as just calling it raw. Their bet: companies will ship 'agent skills' instead of SDKs, teaching AI how to hit your API correctly.
Based on reporting by X — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
There was a time when writing an SDK was the price of admission for getting developers to actually use your API. Nobody wanted to hand-roll auth headers and pagination logic for every service they touched, so companies poured engineering hours into slick client libraries for Python, JavaScript, Go, whatever the flavor of the month happened to be. That calculus has quietly flipped.
The argument now coming out of teams building with AI coding assistants is blunt: writing an SDK wrapper today often costs about the same as just calling the HTTP API directly, because the AI can generate the boilerplate on demand. Ask a model to hit an endpoint, parse the response, and handle retries, and it does it in seconds, no vendor package required. That used to be a day of grunt work. Now it's a prompt.
And if the labor savings of an SDK evaporate, so does most of its reason to exist. Vendor packages come with real costs that developers have always grumbled about: version pinning headaches, dependency bloat, security patching, breaking changes that show up at the worst possible time. If an AI can call your raw API just as easily as it can call your wrapped one, why would anyone volunteer for that maintenance tax?
The proposed replacement is what's being called an agent skill — not a package to install, but a set of instructions that teaches an AI model how to correctly talk to an API. Think of it less like a library and more like documentation written specifically for a machine reader instead of a human one: the right endpoints, the right auth flow, the common pitfalls, all laid out so an agent can call the API correctly without anyone shipping code that has to be versioned and maintained.
This is a small idea with a big implication. If it takes hold, API providers stop maintaining ten SDKs across ten languages and start maintaining one skill file that any capable agent can read. That's a lot less surface area to break, and a lot fewer excuses for outdated documentation drifting from what the API actually does.
My take — AI-written commentary, not fact-checked reporting
I've watched enough SDK version conflicts ruin an afternoon to want this idea to be true, though I'd bet the transition is messier than 'ship a skill file and done' — plenty of APIs are gnarly enough that agents will still hallucinate parameters without a human-tested wrapper somewhere behind the curtain. Still, the direction is right: less code you didn't write cluttering your dependency tree is almost always a win, and this is the first argument for it that doesn't rely on hype.
Read more about this at: X
