Pig.dev’s pivot reveals deeper challenges in building agentic tech for long-term computer use
The Promise: AI Agents for Windows Automation
Pig.dev, a participant in Y Combinator’s Winter 2025 batch, started with an ambitious goal: building AI agents that could control Microsoft Windows desktops — an essential step toward fully autonomous digital assistants for enterprise use.
- This idea targeted a major pain point: automating legacy applications and workflows on Windows systems.
- Pig aimed to be the “Browser Use for Windows”, similar to how Browser Use (another YC alum) helps AI agents interact with web interfaces by translating page elements into a structured format.
The concept had clear appeal. On a recent Y Combinator podcast, partners like Tom Blomfield and David Lieb, along with Replit CEO Amjad Masad, emphasized how crucial desktop and browser automation are to unlocking long-term, productive AI agents.
“The moment that technology works, those two companies are going to do really, really well.” — Amjad Masad
The Reality: A Lack of Market Fit
Despite its technical promise, Pig.dev abandoned its original mission in May 2025. Founder Erik Dunteman revealed that multiple go-to-market strategies failed to gain traction.
- Initially launched as a cloud API product, the idea failed to resonate with users.
- A pivot to a developer tool also didn’t stick.
- Dunteman discovered that most potential customers didn’t want tools — they wanted turnkey automations delivered like consulting services.
But building one-off solutions wasn’t his goal. Dunteman wanted to create scalable developer infrastructure, not act as a service provider.
The Pivot: From Pig.dev to Muscle Mem
Frustrated by the lack of product-market fit, Pig.dev pivoted to a new venture: Muscle Mem, a caching layer for AI agents.
- Muscle Mem helps agents “remember” repeatable tasks, freeing them to focus on reasoning-heavy or novel challenges.
- While no longer focused specifically on Windows, the new project tackles a related bottleneck in AI usability.
“What we’re working on now is directly inspired by and applicable to computer use, just at the developer tooling layer,” Dunteman told TechCrunch.
The Bigger Picture: Why Windows Automation Is So Hard
Dunteman’s experience highlights a critical challenge: agentic technology still struggles with extended use cases and high reliability.
- Agents degrade in accuracy over long sessions as context windows fill up and costs rise.
- Users in traditional industries often seek done-for-you automation, not DIY tools.
That’s not to say the space is dead. On the contrary, companies like Microsoft are actively pushing the frontier:
- In April 2025, Microsoft introduced GUI automation tools for Copilot Studio, available in research preview.
- In July, it announced an agentic feature in Windows 11 to help users manage system settings.
These advances show that Windows automation remains a high-priority frontier — just one that may be best tackled by deep-pocketed incumbents, not early-stage startups.
Lessons from Pig.dev’s Journey
The pivot of Pig.dev underscores a recurring theme in AI development today: bold ideas often run into practical bottlenecks.
- Customer expectations and monetization models matter just as much as technical novelty.
- Infrastructure gaps — such as memory, context handling, and UI complexity — still hinder reliable long-term agents.
- The future of agentic computing may lie in composable systems like Muscle Mem, which chip away at the problem from adjacent angles.
Pig.dev may have changed direction, but its story is a valuable case study in how visionary AI tools meet real-world friction.
