REA computer-use environments
Agents using real Linux desktop apps, graded by their files
Real Linux desktop apps as agentenv computer environments built by REA. Agents use screen, mouse and keyboard; files are graded with no model.
Agents using real Linux desktop apps, graded by their files
Note Gallery: watch agents work in each app
Note Built by Claude Code from a one-line request
Note Built by Claude Code from a one-line request
Note Built by Claude Code from a one-line request in ~20 min; witness held first try; 3/3 agent runs passed.
Note Built by Claude Code from a one-line request in ~24 min; witness held first try; 3/3 agent runs passed.
Note Built by Claude Code from a one-line request in ~20 min; app data, no document; witness held first try; 3/3 passed.
Note Built by Claude Code from a one-line request in ~32 min; witness held first try; 3/3 agent runs passed.
Note Built by Claude Code from a one-line request in ~36 min; witness held first try; 2/2 agent runs passed.
Note Built by Claude Code from a one-line request in ~36 min (hold-out app); witness held first try; 3/3 agent runs passed.
Note Built by Claude Code from a one-line request in ~16 min (hold-out app); witness held first try; 3/3 agent runs passed.
Note Built by Claude Code from a one-line request in ~27 min (hold-out app); witness held first try; 3/3 agent runs passed.
Note Built by Claude Code from a one-line request in ~24 min (hold-out app); witness held first try; 3/3 agent runs passed.
Note Built by Claude Code from a one-line request in ~28 min (hold-out app); witness held first try; 3/3 agent runs passed.
Note Built by Claude Code from a one-line request in ~17 min (third batch); witness held; 3/3 agent runs passed; graded once the app's files settle.
Note Built by Claude Code from a one-line request in ~22 min (third batch); witness held; 2/2 agent runs passed; gzip document, child-text selectors.
Note Built by Claude Code from a one-line request in ~27 min (third batch); witness held; 3/3 agent runs passed; openjdk-17-jre installed beside it.
Note Built by Claude Code from a one-line request in ~28 min (third batch); witness held; 3/3 agent runs passed; librsvg2-common installed beside it.
Note Built by Claude Code from a one-line request in ~30 min (third batch); witness held; 3/3 agent runs passed; gzip book, transaction found by its description.
Note Built by Claude Code from a one-line request in ~18 min (fourth batch); witness held; 3/3 agent runs passed; gzip notebook, note count checked.
Note Built by Claude Code from a one-line request in ~30 min (fourth batch); witness held; 3/3 agent runs passed; gzip label; REA's drafted check kept as is.
Note Built by Claude Code from a one-line request in ~39 min (fourth batch); witness held; 3/3 agent runs passed; slide title found by its placeholder.
Note Built by Claude Code from a one-line request in ~26 min (fifth batch); witness held first try; 3/3 agent runs passed; every check drafted by REA.
Note Built by Claude Code from a one-line request in ~26 min (fifth batch); witness held first try; 3/3 agent runs passed; REA's drafted checks kept as is.
Note Built by Claude Code from a one-line request in ~26 min (fifth batch); witness held first try; 3/3 agent runs passed; rename paired by object id.
Note Built by Claude Code from a one-line request in ~28 min (fifth batch); witness held first try; 2/2 agent runs passed on the final prompt.
Note Built by Claude Code from a one-line request in ~33 min (fifth batch); witness held first try; 2/2 agent runs passed; gzip finance file.
Note Built by Claude Code from a one-line request in ~30 min (sixth batch); witness held first try; 3/3 agent runs passed; tempo checked; ALSA null device seeded.
Note Built by Claude Code from a one-line request in ~27 min (sixth batch); witness held first try; 3/3 agent runs passed; gzip project; every check REA's draft, unedited.
Note Built by Claude Code from a one-line request in ~29 min (sixth batch); witness held first try; 3/3 agent runs passed; element library installed beside it.
Note Built by Claude Code from a one-line request in ~28 min (sixth batch); witness held first try; 3/3 agent runs passed; UML model; class attribute checked by name.
Note Built by Claude Code from a one-line request in ~19 min (sixth batch); witness held first try; 3/3 agent runs passed; layer rename paired by id, with a layer count.