Why a Y Combinator startup tackling AI agents for Windows gave up and pivoted

A startup called Pig.dev that participated in Y Combinator’s Winter 2025 batch was working on a potentially revolutionary idea: AI agentic tech to control a Microsoft Windows desktop.

But in May, the founder announced he was abandoning the tech and pivoting his company to something entirely different: Muscle Mem, a cache system for AI agents that allows them to offload repeatable tasks. 

An early-stage YC company pivoting is nothing out of the ordinary, of course. What’s interesting — and what sparked a dynamic conversation on Thursday’s Y Combinator podcast — is that Pig was working on computer use, one of the big areas that needs to be solved for agents to be truly useful in the workforce. Another company — and another YC alum — that is tackling that for the browser is called Browser Use.

Browser Use surged to popularity when the Chinese agentic tool Manus, which relied on it, went viral. Browser Use essentially scans the buttons and elements of a website to turn them into a more digestible, “text-like” format for agents, helping the AI understand how to navigate and use the website.

During the Y Combinator podcast, released Thursday, partner Tom Blomfield likened Pig to the Browser Use for Windows desktops. The podcast featured Amjad Masad, the founder and CEO of popular vibe-coding startup Replit.

Masad, Blomfield, and YC partner David Lieb were discussing how long-term computer use of hours, rather than minutes, was still a stumbling block for agents. As the context window for reasoning grows, an agent’s accuracy wavers while LLM costs increase.

“The advice I would give founders today is taking either Browser Use or Windows automation with Pig and trying to apply that into enterprise, into a vertical industry,” Blomfield suggested. 

Masad agreed. “The moment that technology works, those two companies are going to do really, really well,” he said.

But alas, Pig’s founder Erik Dunteman has already given up on the idea. In his post in May, he explained that he at first wanted to run a cloud API product (a common way of delivering AI tech). But his customers didn’t want that. So he tried selling it as a dev tool. And they didn’t want that either. 

“What users in the legacy app automation space actually want is to hand me money, and receive an automation,” he said. Essentially, they wanted to hire a consultant to make their desired Windows robotic process automations work for them. 

But Dunteman didn’t want to do one-off projects. He wanted to build development tools. So he abandoned Pig and started working on an AI caching tool. Dunteman declined further comment about his decision to ditch Windows automation, although the Pig.dev website and GitHub documents remain available. 

However, Dunteman did tell us his new tool was inspired by the computer use problem. It is chipping away at it from another angle. The idea is to allow the agent to offload repeated tasks to the Muscle Mem service so the agent can focus on reasoning for new problems and edge cases.

“What we’re working on now is directly inspired by and applicable to computer use, just at the developer tooling layer. I remain very optimistic for computer use as ‘the last mile,’” he told TechCrunch.

That’s not to say that no one is working on Windows automation.

Probably the company furthest along on that is Microsoft. For instance, in April, Microsoft announced it added computer use tech to Copilot Studio for graphical user interfaces like Windows. That tech was released as a research preview.  Plus, earlier this month, Microsoft announced an agentic tool in Windows 11 that helps end users manage settings.