What caught my eye is not the “we built an SDK” part. It’s the way Cursor is positioning agent work as something that should live inside the same boring machinery as the rest of software: source control, reviews, tests, deploys, policies. That’s the right instinct. Most agent demos still feel like a chat wrapper with a few tools bolted on. This reads more like an attempt to make agents behave like services a team can actually own.
I’m still a little skeptical of the neatness of it all, though. “Define it in code, ship it with one command” sounds great, but the hard part is never the packaging. It’s the judgment calls: what the agent is allowed to do, how often it makes weird decisions, and who pays attention when it’s wrong. Cursor seems to know that, which is why the emphasis on evaluations and event replay matters more to me than the deploy story. If those tools are good, this could be genuinely useful. If they’re thin, the rest is just a nicer wrapper around the same brittle behavior.
The part about workflow-specific agents is the strongest argument here. A PR reviewer for one team’s codebase is not the same thing as a PR reviewer for another team’s codebase, and generic “AI reviewer” products usually flatten that difference away. The example of the internal knowledge-base reviewer is exactly the kind of thing I’d expect agents to be good at: narrow, policy-driven, repetitive, and grounded in local context. That feels more credible than broad claims about replacing humans in the loop. I’d actually want to try something like that before I’d trust an agent with anything more ambitious.
I also noticed the subtle shift in how Cursor talks about agents. They’re not just assistants sitting inside the editor anymore; they’re something you deploy onto Cursor’s infrastructure, or even your own Node setup, with their own lifecycle. That’s interesting because it turns Cursor from a product into a platform. That can be powerful, but it also raises the usual platform questions: how portable is this really, and how much of the “SDK” is an invitation to build on Cursor’s terms?
And then there’s the bit about Cursor using this internally for Bugbot, Security Reviewer, and Rollouts. That’s the strongest sales pitch they could make, but also the one I’d treat most cautiously. Internal dogfooding can prove the framework works for Cursor’s own environment. It doesn’t automatically prove it will fit everyone else’s mess. Still, if they’re serious about letting teams define prompts, tools, and tests together, that’s a much more honest direction than pretending agents should be magic.
Reference: Agent SDK のご紹介 · Cursor
