README.md

metadata

license: apache-2.0
pretty_name: AbiWord computer-use environment from a one-line request (REA)
tags:
  - computer-use
  - cua
  - gui-agent
  - rl-environment
  - agentenv
  - desktop
  - linux
  - rea
  - word-processor
  - office
  - from-request

AbiWord: a computer-use environment built from a one-line request

The real AbiWord 3.0 word processor running on an Ubuntu 24.04 desktop, as an agentenv computer environment. An agent completes the task only through screenshots, mouse and keyboard, and a verifier grades the saved files with no model.

Claude Code built this environment from a one-line request in about 13 minutes, with only REA's build skill and MCP tools; How it was made has the request and the steps. AbiWord starts with a blank document. The agent has to open the memo in Documents/book-sale-memo.abw, move the planning meeting from 10:00 AM to 2:30 PM, and save, leaving the memo's other 10:00 AM, an agenda deadline, alone. It sees only the screen and acts only with the mouse and keyboard.

Watch

• Agent runs (Claude Code with Claude Sonnet 5.5): run 1, run 2, run 3
• Witness, which proves the task can be done through the computer: witness.mp4

Each video is the computer's own screen recording at 10 frames per second.

The task

The book sale planning meeting has moved. Change the meeting time in the memo Documents/book-sale-memo.abw to 2:30 PM and save it.

bundle/artifacts/change-meeting-time-check/verify.py reads the app's files through the computer's data/get view and passes only if every check holds:

Check	Kind	File
meeting-time	text_equals	Documents/book-sale-memo.abw
agenda-deadline-kept	text_contains	Documents/book-sale-memo.abw

An .abw file is XML that declares a DOCTYPE; REA reads it without fetching the DTD and finds each paragraph by its xid, never by position. The checks compare a paragraph's whole text, with whitespace collapsed, however AbiWord splits it into <c> runs. The meeting paragraph must read "...at 2:30 PM in the Community Room on the second floor." in full, and the next paragraph must still contain "by 10:00 AM on Wednesday". That second 10:00 AM is there on purpose, so replacing every 10:00 AM fails.

Results

3 of 3 runs passed, about $1.09 per run.

Run	Outcome	Turns	Cost	Video
1	passed	19	$1.20	video
2	passed	18	$1.62	video
3	passed	20	$0.46	video

Run it

You need Docker. agentenv-framework builds the computer: Ubuntu 24.04 with the app installed from Ubuntu's archive (abiword 3.0.5).

pip install agentenv-framework==0.9.1431
hf download earakely-scale/AbiWordFromRequestComputerEnv-REA --repo-type dataset --local-dir abiword-request
agent-env run abiword-request/bundle --task change-meeting-time --agent claude-code --model <model>

The agent needs a model provider configured for agentenv. --task smoke only boots the computer, with no agent. bundle/README.md describes the bundle in detail.

How it was made

Claude Code built this environment from a one-line request, given only REA's build skill and REA's MCP tools:

Build me an AbiWord environment where the agent changes the meeting time in a memo and saves it.

It asked only the skill's upfront questions, wrote no app-specific scripts, and finished in about 13 minutes.

This is the second build of the request. The first, earlier the same evening, exposed a gap in REA: AbiWord's .abw declares a DOCTYPE, which REA refused, so the file could only be checked as raw text lines. AbiWord also splits an edited paragraph into <c> runs, so those line checks failed one correct solve, and that build passed 2 of 3. REA then gained two things: reading XML with a plain DOCTYPE, never fetching the DTD, and checking an element's whole text (text_equals, text_contains) however an app splits it into runs. This rebuild uses them. The steps were REA's usual ones:

1. Capture. REA launched AbiWord on a Linux desktop and recorded its package, windows and opened files while the edit was made. It also recorded the app's state folders before and after.
2. Export. REA turned the capture into an agentenv computer_app. It installs the app, seeds the starting files, starts the app, and waits for its main window. It also serves typed views of the app's files for verifiers.
3. Task. The task's checks were validated against the capture: they fail on the starting state and pass on the captured end state.
4. Witness. In a fresh computer, the edit was made through computer actions alone, and the checks were read at four stages. They held on the first try:

Stage	Checks
seed	fail, as expected
after_edit	pass: the saved memo split the edited paragraph across several <c> runs, and the check still read its whole text
near_miss	fail, as expected: the time was changed to "2:30 AM"
after_reset	fail, as expected: data/reset restored the seed

The agent runs came after the build: Claude Code with Claude Sonnet 5.5 (--model anthropic/claude-sonnet-5-5 --no-emulate, on an amd64 host) solved the task 3 times, each run recorded.

What's inside

bundle/      the agentenv bundle: envs/ (app module, seed files), tasks/, artifacts/ (verifier), evals/
media/       witness and agent-run recordings (MP4) with short GIF previews
results.json the runs above, with each check's result

Limits

• AbiWord starts with a blank document, not the memo, so the agent must open Documents/book-sale-memo.abw itself.
• The prompt fixes the spelling "2:30 PM", and the check compares the whole sentence, so "2:30 p.m." or "14:30" fails.
• The computer installs whatever abiword Ubuntu 24.04's archive holds at build time (3.0.5 for these runs); it is not pinned.
• AbiWord stamps the author user, the computer's generic account, into the memo's metadata.
• Three runs is a small sample: the solve rate shows the task is solvable, not how hard it is.

Licences

The bundle code, the task and the synthetic seed files are Apache-2.0. AbiWord itself is not included: the computer installs it from Ubuntu's archive (abiword 3.0.5) at build time, under its own licence (GPL-2.0-or-later).