Show HN: Coasty.ai – We just hit SOTA on OSWorld (computer-use agent benchmark)
3 comments
You just submitted 4 hours ago: https://news.ycombinator.com/item?id=47115411
New accounts spamming the same submission and having sockpuppets zero-karma users upvoting them: flagged.
Please do visit us @ https://coasty.ai
We've been heads down building Coasty for the past few months and wanted to share a milestone, we just achieved state-of-the-art results on OSWorld, the most comprehensive benchmark for computer-use agents. For those unfamiliar, OSWorld tests whether AI agents can actually use a computer the way a human does navigating real desktop environments, clicking through UIs, filling out forms, managing files across apps.
We're building an agent execution platform for computer-use AI. Think of it like Vercel but for AI agents that operate full desktop environments. You give Coasty a task, and it spins up a sandboxed environment where an agent can see the screen, move the mouse, type, and complete end-to-end workflows things like filing expenses, canceling subscriptions, filling out insurance forms, or navigating legacy enterprise software. Why this matters
Most "AI automation" today is really just API glue between apps that have APIs. But the vast majority of business workflows still live in UIs that were designed for humans, healthcare portals, government forms, internal tools that will never get an API. Computer-use agents unlock automation for all of that. What's next
We're opening up early access. If you're dealing with painful, repetitive computer workflows (especially in healthcare, insurance, legal, or ops), we'd love to talk.
Happy to answer questions about our approach, the benchmark, or anything else.