GodHands: agents that memorize a task instead of re-thinking it every time

Exploring agentic AI, I keep hitting the same wall: an agent that nails a task in the demo and then breaks on the third run. A button moved, a popup appeared, the model re-read the screen and guessed wrong. So a tiny startup attacking exactly that problem caught my attention this week.

GodHands (two founders, Y Combinator) builds a deterministic computer use layer for agents. The pitch is plain: let an agent operate any browser or desktop app, including legacy or internal software with no API, without a custom integration for each one. If a human can see it and click it, the agent can drive it.

The part I find clever is where they put the intelligence. Most computer use today leans on a vision model that re-figures-out the whole screen on every step. That is slow, fragile, and a place where the model can hallucinate an action. GodHands instead separates planning from execution: it learns a task the first time, then replays it deterministically and fast, while still adapting to the unexpected (a random popup, a crash). It is roughly how we work ourselves: clumsy with a new app, quick once we know it. Every action becomes repeatable, which pulls hallucination out of the execution layer, and they keep an immutable log of each step for debugging and compliance.

What strikes me as a frontend dev is the reframing. The hard problem for agents in real workflows is not capability, it is reliability, and reliability is an infrastructure problem, not a smarter-prompt problem.

If this approach holds, the boring enterprise apps no one wanted to integrate suddenly become automatable. That is the unglamorous work that actually moves things.

Sources

• GodHands (official site): https://godhands.dev/
• Y Combinator, GodHands: https://www.ycombinator.com/companies/godhands