Half a Million Computer Actions in Seven Days
Half a Million Computer Actions in Seven Days
The Terminator desktop automation library hit half a million computer actions in its first week of heavy usage. Clicks, keystrokes, scrolls, window manipulations, text selections - each one logged, each one analyzed. The numbers revealed patterns we did not expect.
What 500K Actions Look Like
Most actions are mundane. Click a button. Type a character. Scroll down. These account for about 80 percent of total volume. The remaining 20 percent are the complex multi-step sequences - drag and drop, modifier-key combinations, precise coordinate targeting - that take disproportionate engineering effort.
The failure rate distribution is even more skewed. Simple clicks fail less than 0.1 percent of the time. Complex drag operations fail about 8 percent of the time. The lesson: simple actions at scale are reliable. Complex actions at scale need fallback strategies.
Speed vs Accuracy
At low volume, you can afford to verify every action. Take a screenshot after each click. Confirm the expected element appeared. At 500,000 actions per week, verification overhead dominates. You cannot screenshot after every action without turning a 100ms operation into a 2-second operation.
The solution is sampled verification. Verify a percentage of actions. Increase the verification rate for action types with higher failure rates. Skip verification entirely for actions with near-zero failure rates. This keeps throughput high while catching failures before they cascade.
What This Means for Desktop Agents
Desktop agents are only as good as their action layer. A brilliant planning model paired with an unreliable action layer produces frustration. The 500K action dataset gives us failure mode data that makes the action layer predictable. When the agent clicks, it knows the probability of success before it acts.
Reliability at scale is the unglamorous foundation that makes intelligent automation possible.
Fazm is an open source macOS AI agent. Open source on GitHub.