From Sand to Superintelligence · Drill cards · Chapter 38
Drills
The Browser Becomes the Worker
9 atomic recall cards. Export to Anki and let spaced repetition do its slow work.
In Anki: File → Import, choose this TSV, set field separator to Tab, deck = Sand to Silicon · Ch 38, note type = Basic.
| Front | Back |
|---|---|
| When did Anthropic ship Claude with computer-use capability? | October 2024; OpenAI's Operator followed. |
| What is the median latency per browser-action step? | ~5–15 seconds end-to-end per step. |
| What WebArena benchmark completion rate did the best agents achieve in late 2025? | 50–70% of tasks. |
| Name the three perception modalities a reliable computer-use agent fuses. | Rendered screenshot, DOM tree, and accessibility tree. |
| What does the agent inherit by acting as if it were a user in the browser? | Whatever permissions the user it is acting as has — convenient for capability, dangerous for security. |
| Name the three characteristic failure modes of computer-use agents. | Drift (diverging from the goal), loops (retrying a failing action), and confabulated success (claiming completion without actually submitting). |
| What is the economic case for using an agent when it is slower than a human? | Unit cost and parallelism — one agent can run a thousand tasks simultaneously for less than the cost of a single human doing them in series. |
| What does the chapter say is 'genuinely lost' as agents replace human users of UIs? | The ambient telemetry — the subtle signals about user intent that layout and forms could elicit from a human filling them in. |
| What kind of companies will be most harmed by the computer-use shift, per the chapter? | Companies whose moat was a hard-to-use UI that customers tolerated — an agent does not tolerate, it just leaves. |