Coasty wants to kill RPA by letting agents just use the computer
While the big labs fight over foundation models, a YC company named Coasty is betting on the boring, lucrative middle: getting an agent to actually operate the software companies already run. Not API calls — the desktop. It drives web apps, old Windows tools, mainframe terminals, internal portals, and Citrix sessions by sight and click, the way a human does. When a button moves or a page gets redesigned, it re-reads the screen and keeps going. Nothing to rewrite.
That last part is the whole pitch against RPA. Traditional robotic process automation breaks the moment a selector changes, which is why enterprises spend fortunes maintaining brittle scripts. A computer-use agent that survives UI updates and just follows the SOP is a different economic animal. Coasty claims the top spot on OSWorld at around 85%, against numbers in the 30s and 40s for general-purpose agents from the big labs — worth treating as a vendor benchmark until independently checked, but the gap they are pointing at is real.
The reason legacy is the smart wedge: it is where the money and the pain both live. Banks, insurers, and hospitals run decades-old systems with no API and no vendor willing to touch them. RPA sort of works and everyone hates it. If an agent can reliably drive a green-screen mainframe by looking at it, that is a market measured in seats today, not a demo.
The risk is reliability at scale — computer use is still the hardest agent surface, and 85% on a benchmark is not 99.9% on a payroll run where a wrong click costs real money. But the framing is sharp: the winning agent might not be the smartest one, it might be the one that shows up on the screen your business already depends on. Launched on Product Hunt this week; see https://coasty.ai/
← Back to all articles
That last part is the whole pitch against RPA. Traditional robotic process automation breaks the moment a selector changes, which is why enterprises spend fortunes maintaining brittle scripts. A computer-use agent that survives UI updates and just follows the SOP is a different economic animal. Coasty claims the top spot on OSWorld at around 85%, against numbers in the 30s and 40s for general-purpose agents from the big labs — worth treating as a vendor benchmark until independently checked, but the gap they are pointing at is real.
The reason legacy is the smart wedge: it is where the money and the pain both live. Banks, insurers, and hospitals run decades-old systems with no API and no vendor willing to touch them. RPA sort of works and everyone hates it. If an agent can reliably drive a green-screen mainframe by looking at it, that is a market measured in seats today, not a demo.
The risk is reliability at scale — computer use is still the hardest agent surface, and 85% on a benchmark is not 99.9% on a payroll run where a wrong click costs real money. But the framing is sharp: the winning agent might not be the smartest one, it might be the one that shows up on the screen your business already depends on. Launched on Product Hunt this week; see https://coasty.ai/
Comments